<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Sentinelleai</title>
    <description>The latest articles on DEV Community by Sentinelleai (@sentinelleai).</description>
    <link>https://dev.to/sentinelleai</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4213374%2F7880cb0e-6e2d-4ffe-9d58-98ec64330a25.jpg</url>
      <title>DEV Community: Sentinelleai</title>
      <link>https://dev.to/sentinelleai</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/sentinelleai"/>
    <language>en</language>
    <item>
      <title>Your backup server is the crown jewel: why ransomware crews go after backups first</title>
      <dc:creator>Sentinelleai</dc:creator>
      <pubDate>Thu, 12 Nov 2026 16:19:00 +0000</pubDate>
      <link>https://dev.to/sentinelleai/your-backup-server-is-the-crown-jewel-why-ransomware-crews-go-after-backups-first-3d32</link>
      <guid>https://dev.to/sentinelleai/your-backup-server-is-the-crown-jewel-why-ransomware-crews-go-after-backups-first-3d32</guid>
      <description>&lt;p&gt;Between 10 and 19 September 2026, an attacker worked its way through a UK school estate. It did not start with the school's data. It started with the &lt;strong&gt;backup servers&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;According to &lt;a href="https://www.bitdefender.com/en-us/blog/businessinsights/3am-ransomware-hybrid-ai-attack" rel="noopener noreferrer"&gt;Bitdefender's report&lt;/a&gt;, published on 3 November and &lt;a href="https://www.bleepingcomputer.com/news/security/uk-school-estate-hit-by-3am-ransomware-in-suspected-ai-assisted-attack/" rel="noopener noreferrer"&gt;covered by BleepingComputer&lt;/a&gt;, an affiliate of the 3AM ransomware group used a modified public PowerShell script to read stored credentials from &lt;strong&gt;three Veeam backup servers&lt;/strong&gt; and decrypt them. Nine days after that first step, about 37 machines were encrypted.&lt;/p&gt;

&lt;p&gt;Two things make this case worth reading. The first is the order of events: backups first, encryption last. The second is a suspicion, which we label as one: Bitdefender thinks parts of the intrusion may have been driven by an AI model in an automated loop. It could not confirm that. We come back to it below. The lesson holds either way.&lt;/p&gt;

&lt;p&gt;One point before anything else. Nothing in the report says Veeam, or any backup product, was vulnerable. The attacker already controlled the servers, and it read credentials that the backup software has to keep in order to do its job. That is the point: a backup system must hold the keys to everything it protects, so whoever reaches it gets those keys.&lt;/p&gt;

&lt;h2&gt;
  
  
  What happened, in order
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdp13c6kde718v9mm120k.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdp13c6kde718v9mm120k.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Bitdefender describes a school estate of 47 systems, about 21 of them servers (8 domain controllers), split across several separate Active Directory domains. Ten workstations still ran Windows 10 and one domain controller ran Windows Server 2012 R2. How the attacker got in is &lt;strong&gt;unknown&lt;/strong&gt;. Bitdefender says so plainly.&lt;/p&gt;

&lt;p&gt;What it could see started on an internal machine with no Bitdefender agent on it. Every attacker action Bitdefender observed came from that &lt;strong&gt;unmanaged host&lt;/strong&gt;, and Bitdefender could not examine it.&lt;/p&gt;

&lt;p&gt;From there, according to Bitdefender:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;10 to 11 September: backup credentials.&lt;/strong&gt; A modified public script (Veeam-Get-Creds.ps1) read stored credentials from the Veeam databases on three backup servers and decrypted them. A Chrome credentials file was also copied from a terminal server.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;12 to 14 September: domain admin.&lt;/strong&gt; The attacker checked who the Domain Admins were, downloaded Mimikatz, saved registry hives from the Windows Server 2012 R2 domain controller, and tried to dump LSASS memory (the outcome is unknown). It created a backdoor account named &lt;strong&gt;library&lt;/strong&gt; and added it to Domain Admins on the 14th.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;15 September: a second backdoor and data theft.&lt;/strong&gt; A second account, &lt;strong&gt;svc_bkpmon&lt;/strong&gt;, joined Domain Admins. Bitdefender notes the names were chosen to blend into a school's account list. About 40 finance files were staged and sent to the unmanaged host.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;18 September: three domains joined.&lt;/strong&gt; The attacker built &lt;strong&gt;two-way trusts between three separate domains&lt;/strong&gt;, then pushed a script to about 40 machines to give its backdoor account access. A trust does not grant permissions by itself, which is why that second step was needed. Bitdefender counts about 50 machines reached from the unmanaged host, slightly more than the 47 in the machine list it describes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;19 September: security tools, backups, encryption.&lt;/strong&gt; Scripts tried to remove Sophos and Bitdefender protection, and &lt;strong&gt;anti-tampering controls blocked them&lt;/strong&gt;. System-state backups were deleted with &lt;code&gt;wbadmin delete systemstatebackup -keepVersions:0&lt;/code&gt;. The 3AM encryptor then ran from 05:07 to 05:33, and about 37 machines were encrypted.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3822r2smlujara53xx8i.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3822r2smlujara53xx8i.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Bitdefender states that nothing in the toolkit was novel: built-in Windows tools plus public tools. That is good news for defenders, because standard controls should stop standard techniques.&lt;/p&gt;

&lt;h2&gt;
  
  
  The suspicion: was an AI model driving?
&lt;/h2&gt;

&lt;p&gt;Bitdefender calls the operation a "hybrid" and is careful about it. On 13 September, on one domain controller, a PowerShell command ran twice, byte for byte, 41 seconds apart. Forty-nine seconds later a third run changed only its last call, and 47 seconds after that the attacker gave up on PowerShell and switched to wmic. Bitdefender says the same rhythm held across nine retries. Other signs it lists: a call to Veeam's REST API on the wrong port and with the wrong authentication scheme (the documented examples use another), Bash-style syntax inside Windows PowerShell, and numbered script revisions.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fook4hw7vkjcfyr1buve0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fook4hw7vkjcfyr1buve0.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Against that, Bitdefender lists the counter-indicators: a conventional encryptor, public tooling, no natural-language reasoning text in the payloads, and no decision cycle faster than about 40 to 60 seconds. Because it could not access the attacker's machine, it cannot say which model was involved, and it cannot rule out a person pasting commands and results between a shell and a chat tool. Of its three claims, it rates "a model generated the commands" as strongly supported, "an agentic loop chose each next action" as behaving like one, and "no human involved in the hands-on phase" as unsupported. It also stresses this is one case.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Treat the AI part as unconfirmed.&lt;/strong&gt; The practical takeaway is Bitdefender's own: a tool that retries every minute will beat a defense that depends on a person correlating alerts quickly. Prevent, rather than race.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why backups come first
&lt;/h2&gt;

&lt;p&gt;A backup server is among the most useful machines on a small network for an attacker, for three reasons:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;It holds credentials.&lt;/strong&gt; To back up servers, it stores accounts with broad access to them. Read them out and you hold many of the keys at once.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It holds the way back.&lt;/strong&gt; Delete or encrypt the backups and the victim cannot restore, so paying becomes the only way out. Here, system-state backups were deleted minutes before the encryptor finished.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It is often trusted.&lt;/strong&gt; It sits inside the production domain, its traffic is expected, and nobody looks at it until a restore is needed.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Protections that exist, tied to what happened
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6y7zulhvnte00kaoemmp.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6y7zulhvnte00kaoemmp.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Keep a copy the domain cannot delete (3-2-1-1-0).&lt;/strong&gt; Veeam's best-practice guidance sets out the 3-2-1-1-0 rule: three copies, on two kinds of media, one off site, one copy "offline, air-gapped, or immutable", and zero errors when you verify a restore. Veeam's community write-up explains that an immutable backup cannot be changed or deleted even by a compromised admin account. Immutability is a category with several implementations: a Hardened Linux Repository (Veeam's own term), S3-compatible object storage with Object Lock, tape, or a disk you unplug. Bitdefender's advice is the same in plain terms: keep immutable recovery copies that no domain account can delete.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Treat the backup server as a target, not as a utility.&lt;/strong&gt; Bitdefender recommends isolating backup servers and scoping backup accounts per domain, and never storing privileged credentials for several domains in one place. In practice: give the backup system its own accounts, do not reuse a domain admin for it, and consider keeping it out of the production domain if your design allows. Turn on MFA for the backup console where the product supports it, and keep service accounts to the minimum rights the job needs.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Make local admin passwords unique with Windows LAPS.&lt;/strong&gt; Microsoft's Windows LAPS rotates each device's local administrator password and stores it for authorized admins, to counter pass-the-hash and lateral movement. It is free, built into supported Windows 11 and Windows Server (Windows Server 2019 and later with the April 2023 update), and covers Active Directory-joined machines.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Alert on new domain admins and new trusts.&lt;/strong&gt; Both were central here, and Windows logs them. Microsoft's reference says event &lt;strong&gt;4720&lt;/strong&gt; fires whenever a user account is created (on domain controllers, member servers and workstations) and that some organizations monitor every one. Event &lt;strong&gt;4706&lt;/strong&gt; records a new trust created to a domain (on domain controllers only), and Microsoft's guidance is that any change to domain trusts should alert, with unplanned ones investigated. Add the group-membership change events for Domain Admins and Enterprise Admins. Bitdefender asks for the same: put trusts and domain-admin accounts under change control and alert on any that appear without it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Watch behaviors, with free tools.&lt;/strong&gt; &lt;a href="https://learn.microsoft.com/en-us/sysinternals/downloads/sysmon" rel="noopener noreferrer"&gt;Sysmon&lt;/a&gt; from Microsoft logs process creation with full command lines (event 1), processes opening other processes such as LSASS (event 10), registry changes (events 12 and 13) and named pipes (events 17 and 18). &lt;a href="https://github.com/SigmaHQ/sigma" rel="noopener noreferrer"&gt;Sigma&lt;/a&gt; is an open, generic format for detection rules over such logs, and &lt;a href="https://documentation.wazuh.com/current/getting-started/index.html" rel="noopener noreferrer"&gt;Wazuh&lt;/a&gt; is an open-source SIEM and XDR platform with log analysis, file integrity monitoring and a Windows agent. Bitdefender suggests alerting on credential-reading access to backup-software databases, and on &lt;code&gt;reg save&lt;/code&gt; of the SECURITY, SYSTEM or SAM hives by anything other than a known backup or admin job.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;6. Turn on tamper protection in your endpoint security.&lt;/strong&gt; In this case it worked: both removal attempts were blocked. It only protects machines that have an agent, which leads to the next point.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;7. Know every device on the network.&lt;/strong&gt; The attacker staged from a machine with no agent. Keep an inventory, put an endpoint agent on everything that can run one, and investigate unknown hosts. Bitdefender also recommends limiting WMI, remote registry, scheduled-task creation and admin shares to known management hosts, and blocking interpreters such as portable Python on machines that never use them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;8. Test the restore.&lt;/strong&gt; The "0" in 3-2-1-1-0 is a drill, not a setting. The Veeam community write-up calls a backup that has never been tested effectively unverified. Restore one server to an isolated place, time it, and write down who does it when the domain is gone.&lt;/p&gt;

&lt;h2&gt;
  
  
  A checklist for tonight
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F26amhsvjq0ej1c4t1p9b.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F26amhsvjq0ej1c4t1p9b.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;List where your backups live. Is at least one copy unreachable by any domain account (immutable, object-locked, or physically offline)?&lt;/li&gt;
&lt;li&gt;Is the backup server in the production domain? If yes, what can a domain admin do to it?&lt;/li&gt;
&lt;li&gt;Which accounts does the backup software use, and what can they reach? Remove anything beyond what the job needs.&lt;/li&gt;
&lt;li&gt;Is MFA enabled on the backup console?&lt;/li&gt;
&lt;li&gt;Do domain controllers audit account creation, group changes and trust creation, and does someone get an alert? (Events 4720 and 4706 are a start.)&lt;/li&gt;
&lt;li&gt;Are local admin passwords unique per machine (Windows LAPS)?&lt;/li&gt;
&lt;li&gt;Does every machine have an endpoint agent with tamper protection on? List the ones that do not.&lt;/li&gt;
&lt;li&gt;Is a Windows 10 workstation or a Windows Server 2012 R2 machine still running? Plan its replacement.&lt;/li&gt;
&lt;li&gt;Restore one system from backup this month, and time it.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The lesson
&lt;/h2&gt;

&lt;p&gt;The attacker in this case did not break a backup product. It reached servers, read what they had to store, and used it. Each step has a published defense, and none of the defenses is exotic. Keep one copy that no domain account can reach, treat the backup server as one of the most sensitive machines you own, and make sure you hear about a new domain admin or trust before an attacker's work is done.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;I build &lt;a href="https://sentinelleai.app/for/business?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=backup-server" rel="noopener noreferrer"&gt;Sentinelleai&lt;/a&gt;, which watches the systems of small businesses and their IT providers for this kind of attack.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Sources: Bitdefender Business Insights, "3AM Hits European Schools: A Hybrid AI Ransomware Attack, Part Human, Part Agentic Loop" (3 Nov 2026), &lt;a href="https://www.bitdefender.com/en-us/blog/businessinsights/3am-ransomware-hybrid-ai-attack" rel="noopener noreferrer"&gt;https://www.bitdefender.com/en-us/blog/businessinsights/3am-ransomware-hybrid-ai-attack&lt;/a&gt; ; BleepingComputer, "UK school estate hit by 3AM ransomware in suspected AI-assisted attack" (3 Nov 2026), &lt;a href="https://www.bleepingcomputer.com/news/security/uk-school-estate-hit-by-3am-ransomware-in-suspected-ai-assisted-attack/" rel="noopener noreferrer"&gt;https://www.bleepingcomputer.com/news/security/uk-school-estate-hit-by-3am-ransomware-in-suspected-ai-assisted-attack/&lt;/a&gt; ; Veeam Best Practice Guide, &lt;a href="https://bp.veeam.com/vbr/1_Assessment/A_Infra_Overview/infra_overview.html" rel="noopener noreferrer"&gt;https://bp.veeam.com/vbr/1_Assessment/A_Infra_Overview/infra_overview.html&lt;/a&gt; ; Veeam Community, "The 3-2-1-1-0 rule in practice", &lt;a href="https://community.veeam.com/blogs-and-podcasts-57/the-3-2-1-1-0-rule-in-practice-how-to-actually-implement-it-with-veeam-12799" rel="noopener noreferrer"&gt;https://community.veeam.com/blogs-and-podcasts-57/the-3-2-1-1-0-rule-in-practice-how-to-actually-implement-it-with-veeam-12799&lt;/a&gt; ; Microsoft Learn, Windows LAPS overview, &lt;a href="https://learn.microsoft.com/en-us/windows-server/identity/laps/laps-overview" rel="noopener noreferrer"&gt;https://learn.microsoft.com/en-us/windows-server/identity/laps/laps-overview&lt;/a&gt; ; Microsoft Learn, security auditing reference for events 4706 and 4720 ; Microsoft Learn, Sysmon, &lt;a href="https://learn.microsoft.com/en-us/sysinternals/downloads/sysmon" rel="noopener noreferrer"&gt;https://learn.microsoft.com/en-us/sysinternals/downloads/sysmon&lt;/a&gt; ; Sigma, &lt;a href="https://github.com/SigmaHQ/sigma" rel="noopener noreferrer"&gt;https://github.com/SigmaHQ/sigma&lt;/a&gt; ; Wazuh documentation, &lt;a href="https://documentation.wazuh.com" rel="noopener noreferrer"&gt;https://documentation.wazuh.com&lt;/a&gt;. Facts as of 11 Nov 2026; Bitdefender's analysis may be updated.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>devops</category>
      <category>windows</category>
      <category>cybersecurity</category>
    </item>
    <item>
      <title>How to find bug bounty and disclosure programs, and get a report accepted</title>
      <dc:creator>Sentinelleai</dc:creator>
      <pubDate>Sun, 08 Nov 2026 23:46:50 +0000</pubDate>
      <link>https://dev.to/sentinelleai/how-to-find-bug-bounty-and-disclosure-programs-and-get-a-report-accepted-4h6l</link>
      <guid>https://dev.to/sentinelleai/how-to-find-bug-bounty-and-disclosure-programs-and-get-a-report-accepted-4h6l</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://sentinelleai.app/guides/how-to-find-bug-bounty-programs" rel="noopener noreferrer"&gt;sentinelleai.app&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuwi2ikne480bf2xz4zbe.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuwi2ikne480bf2xz4zbe.png" alt="A magnifying glass over a map of program sources: VDP, bounty programs, Web3, security.txt and government pages" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Programs are published in many places. The skill is knowing where to look, and reading the rules before you touch anything.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;If you search for "responsible disclosure bounty" or "bug bounty", you are probably trying to answer two questions. Where can I report a problem I found, and is there any reward for it? This guide covers both, in the order you should do them: understand what a program promises, find the right one, read its policy, and only then test and report.&lt;/p&gt;

&lt;p&gt;Platforms change their pages from time to time, so treat every name below as a starting point and trust the live policy page over any summary, including this one.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Bounty, disclosure program, coordinated disclosure
&lt;/h2&gt;

&lt;p&gt;A &lt;strong&gt;vulnerability disclosure program (VDP)&lt;/strong&gt; tells outsiders how and where to report a security problem. HackerOne's help centre describes it this way: a VDP "solely gives clear guidelines for how an organization would like to be notified of potential security vulnerabilities found by external third parties." It typically lists a scope, a process, and a safe-harbour statement, which HackerOne defines as "a statement that assures finders that they won't be penalized or have legal action taken against for the vulnerabilities that they find."&lt;/p&gt;

&lt;p&gt;A &lt;strong&gt;bug bounty program&lt;/strong&gt; adds money. In HackerOne's words, it "incentivizes external third parties to find security vulnerabilities in a company's software and report them directly to the company," and the finders "are rewarded with monetary prizes." HackerOne also notes the usual difference in incentive: bounties pay, while VDPs "typically offer thanks and recognition."&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Coordinated disclosure&lt;/strong&gt; is the practice around both: you report privately, the owner fixes, and details become public on an agreed schedule. Section 7 covers the timing.&lt;/p&gt;

&lt;p&gt;The label matters less than the policy, which decides what you may do. Some organisations publish a VDP with no reward, some run a bounty with a narrow scope, and some publish nothing.&lt;/p&gt;

&lt;h3&gt;
  
  
  The legal line
&lt;/h3&gt;

&lt;p&gt;Reading the policy first is what separates research from intrusion. A published scope and a safe-harbour clause are the owner's written statement of what they accept. Without them, you are guessing.&lt;/p&gt;

&lt;p&gt;Government policy can help, but it is not a licence. The US Department of Justice's charging policy for the Computer Fraud and Abuse Act says prosecutors "should decline prosecution if available evidence shows the defendant's conduct consisted of, and the defendant intended, good-faith security research." It defines that as accessing a computer "solely for purposes of good-faith testing, investigation, and/or correction of a security flaw," carried out "in a manner designed to avoid any harm to individuals or the public." The policy says it creates no enforceable right for anyone and covers US federal charging only. Other countries and private civil claims are different matters, so a clear safe-harbour clause from the owner is still what you want.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Where programs are listed
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffglohtxvv07xy82v9hmx.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffglohtxvv07xy82v9hmx.png" alt="Four sources (bounty platforms, Web3 programs, independent lists, company pages) all pointing to the researcher" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Four places to look. All of them end in the same step: reading the policy page.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Bounty platforms.&lt;/strong&gt; HackerOne, Bugcrowd, Intigriti and YesWeHack each publish a directory of programs. Intigriti's public page opens with "Below is a list of public bug bounty programs," and YesWeHack's programs page separates public bug bounty and VDP entries. Private programs are invite-only, so a directory is never the full picture.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Web3 programs and contests.&lt;/strong&gt; Immunefi runs a "bug bounty platform" for crypto projects. Audit contests are a different model and worth separating in your head. Code4rena describes itself as a "competitive audit and bounty platform" where researchers review code "in exchange for bounties provided by sponsoring projects," and sponsors "create prize pools." Sherlock defines its contest as "a public Web3 security review where researchers compete to find valid vulnerabilities in a team's codebase for a prize pool." Cantina also hosts competitions. In a contest you compete inside a fixed window against other researchers, findings are judged, and Sherlock says its process de-duplicates them. In a standing bounty you report to one program on its own timeline. The rules, rewards and risk of a duplicate differ, so read the contest page, not a bounty template.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Independent lists.&lt;/strong&gt; disclose.io is an open project that, in its own words, maintains "open directories of disclosure programs and bug bounty platforms." Its directory lists programs with "scope, policy URLs, and safe-harbor language," and grades them on a maturity scale. At the top end, "Full" means the program "explicitly authorises testing"; at the bottom, it only gives a contact address.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Government VDPs.&lt;/strong&gt; CISA operates a platform for US federal civilian agencies. It launched in July 2021 "to ensure that Federal Civilian Executive Branch (FCEB) agencies benefit from the expertise of the research community," and agencies use it to "collaborate with public security researchers." Other governments run their own schemes, so check the national CERT or cyber agency for the country you are in.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The company's own pages.&lt;/strong&gt; Many organisations publish a security or responsible-disclosure page, and many put a &lt;code&gt;security.txt&lt;/code&gt; file on their site. That is the next section.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. The security.txt file
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F25gyv6j4te0w8yzrbbx7.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F25gyv6j4te0w8yzrbbx7.png" alt="A browser window showing a security.txt file with each field labelled by a short callout" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;An example security.txt. Values are fictional.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;RFC 9116 defines &lt;code&gt;security.txt&lt;/code&gt; as a machine-parsable file that helps organisations describe their disclosure practices, so researchers can more easily report vulnerabilities. For web services, the file lives at &lt;code&gt;/.well-known/security.txt&lt;/code&gt;. The RFC says a copy may also sit at the top-level path "for legacy compatibility," and that if both exist, the one under &lt;code&gt;/.well-known/&lt;/code&gt; "MUST be used." It must be served over HTTPS as plain text.&lt;/p&gt;

&lt;p&gt;The fields, per the RFC:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Contact&lt;/strong&gt; (required): how to report, as a &lt;code&gt;mailto:&lt;/code&gt;, &lt;code&gt;tel:&lt;/code&gt; or &lt;code&gt;https://&lt;/code&gt; URI. Several can be listed, "in order of preference."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Expires&lt;/strong&gt; (required): the date after which the file "is considered stale and should not be used."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Policy&lt;/strong&gt;: a link to the vulnerability disclosure policy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Canonical&lt;/strong&gt;: where the file is meant to be found.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Preferred-Languages&lt;/strong&gt;: languages for reports. If absent, "security researchers may assume that English is the language to be used."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Encryption&lt;/strong&gt;: a link to a key for encrypted reports.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Acknowledgments&lt;/strong&gt;: a page recognising researchers.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hiring&lt;/strong&gt;: security-related jobs.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;How to read one: check that &lt;code&gt;Expires&lt;/code&gt; is still in the future, check that the URL you fetched is listed under &lt;code&gt;Canonical&lt;/code&gt; if that field exists, and follow &lt;code&gt;Policy&lt;/code&gt; before anything else. The RFC warns that a file may be tampered with, and tells researchers to "record any redirects" because they can lead somewhere an attacker controls.&lt;/p&gt;

&lt;p&gt;The most important sentence in the RFC for you is this one: "researchers shouldn't assume that the presence or absence of a 'security.txt' file grants or denies permission for security testing." The file tells you where to send a report. It does not authorise testing. The permission, if there is any, is in the policy.&lt;/p&gt;

&lt;p&gt;A &lt;code&gt;security.txt&lt;/code&gt; applies only to the domain it is served from, "not to any of its subdomains or parent domains."&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Search operators
&lt;/h2&gt;

&lt;p&gt;Search engines can find published policies. Operators narrow the results. Support and behaviour differ between engines and change over time, so test them on the engine you use.&lt;/p&gt;

&lt;p&gt;Patterns researchers use:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;"responsible disclosure" "reward"&lt;/code&gt; finds pages containing both phrases.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;"vulnerability disclosure policy" "safe harbor"&lt;/code&gt; finds policies that mention legal protection.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;inurl:security.txt&lt;/code&gt; or &lt;code&gt;inurl:responsible-disclosure&lt;/code&gt; matches words in the address.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;site:example.com "report a vulnerability"&lt;/code&gt; looks for a policy on one organisation's own domain.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;intext:"bug bounty"&lt;/code&gt; matches the phrase in the page body.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Limits are real. Results are keyword matches, not a verified list, and old pages stay indexed after a program closes. A hit is a lead: open the page, check the date, and confirm the program on the owner's own domain or a platform.&lt;/p&gt;

&lt;p&gt;This technique is for finding the published rules. It is not a way to find systems to test.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. Before you test
&lt;/h2&gt;

&lt;p&gt;Read these parts of the policy, every time:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Scope.&lt;/strong&gt; Which domains, apps, contracts or versions are included. If an asset is not listed, assume it is out.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Out-of-scope list.&lt;/strong&gt; Many programs exclude vulnerability classes (for example, purely theoretical issues) and actions (for example, denial of service or social engineering). Immunefi's rules, for instance, list "automated testing of services that generates significant amounts of traffic" and "any testing with mainnet or public testnet contracts" as prohibited, with the second described as "grounds for an immediate and permanent ban." That is Immunefi's rule; yours will be whatever your program says.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rules of engagement and accounts.&lt;/strong&gt; Rate limits, allowed tools, test environments. Use accounts you created for the purpose, never other users' accounts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Safe harbour.&lt;/strong&gt; Look for explicit wording. "We will not pursue legal action for good-faith research within this policy" is stronger than silence.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Data handling.&lt;/strong&gt; Stop as soon as you can show the issue exists. Immunefi's rules prohibit "unauthorized disclosure or access of sensitive information beyond what is necessary to submit the report." Do not download, keep or share more than you need.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;If you see someone else's data:&lt;/strong&gt; stop, do not read further, do not copy it, record only the minimum needed to show the issue (for example, that a record belonging to another account was returned, not its contents), and say so in the report. Delete what you do hold once the report is resolved, unless the program says otherwise.&lt;/p&gt;

&lt;p&gt;If anything is unclear, ask through the program's contact channel before testing.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. What gets a report accepted
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr7ezeuyhi7dh1u018410.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr7ezeuyhi7dh1u018410.png" alt="A report card with six parts: title, asset, steps, proof, impact and fix" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The six parts a triager looks for.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;A report is accepted when a triager can reproduce it quickly and see why it matters.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;A specific title.&lt;/strong&gt; "Stored XSS in profile bio on app.example.com", not "Critical vulnerability".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The affected asset.&lt;/strong&gt; The exact URL, contract address, app version or endpoint, matching an in-scope entry.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Steps to reproduce.&lt;/strong&gt; Numbered, starting from a clean account or state. Someone else should be able to follow them without guessing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A proof of concept.&lt;/strong&gt; Request and response, a short video, or a script that runs, showing the effect on data you own.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Impact and severity.&lt;/strong&gt; Who is affected and what an attacker gains. Argue the severity from the facts rather than naming a level.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A suggested fix.&lt;/strong&gt; Optional, but it shortens the conversation.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Check for duplicates first: read the program's disclosed reports and known-issues list if it has them. Immunefi prohibits "reporting a bug that has already been publicly disclosed," and Sherlock says its judging process clears duplicates in contests.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why scanner output and AI-written reports are rejected
&lt;/h3&gt;

&lt;p&gt;Platforms say this in writing. Immunefi's published rules list among prohibited behaviour: "Submitting AI-generated/automated scanner bug reports that lack the required information regarding the vulnerability's impact on the reported asset," and "Placeholder bug submissions, i.e., bugs that have a vague title, very few details, and no reproducible steps." They also list "submitting a bug report with no PoC or an incomplete PoC if it is required by the project's bug bounty program," and "misrepresenting severity: claiming that a bug report is critical when it clearly is not."&lt;/p&gt;

&lt;p&gt;Bugcrowd published a policy update on AI-generated submissions. It states that "you, the researcher, are responsible for the accuracy of the submission," and that accounts that submit at least 10 consecutive invalid reports "will be reviewed", with a possible 30-day suspension where the reports are "attributable to automated or AI-generated activity without sufficient validation."&lt;/p&gt;

&lt;p&gt;The pattern is the same everywhere. A tool can help you find a lead. You still have to confirm it by hand, show the impact, and write it up yourself. A raw scanner finding is an unverified claim, and the triager has no reason to believe it.&lt;/p&gt;

&lt;p&gt;Also note where to report. Immunefi lists "submitting bugs via email or any channel other than the Immunefi platform" as prohibited. Use the channel the program names.&lt;/p&gt;

&lt;h2&gt;
  
  
  7. After you report
&lt;/h2&gt;

&lt;p&gt;Expect an automated acknowledgement, then triage, where someone tries to reproduce the issue and may ask questions. Answer promptly and stay factual. Many programs publish target response times; a VDP policy, per HackerOne, may state "the expected time finders should wait between submission and the first response" and "when finders can publicly disclose their found vulnerabilities."&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frx83ur1ze304t6yrrwmp.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frx83ur1ze304t6yrrwmp.png" alt="A timeline from report to patch to publication, with a 90 day bar and a 30 day tail after the patch" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Timings from Google Project Zero's policy. Other programs set their own.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Disclosure windows.&lt;/strong&gt; A well-known reference is Google Project Zero's "90+30" policy: "a vendor has 90 days after Project Zero notifies them about a security vulnerability to make a patch available to users. If they make a patch available within 90 days, Project Zero will publicly disclose details of the vulnerability 30 days after the patch has been made available." Its worked example: patched at day 47, details public on day 77. If the vendor cannot patch in 90 days, details are published at the end of that period, and the policy allows a grace period in some cases. Actively exploited issues follow a 7-day clock. This is one organisation's policy, not a universal rule. Your program may forbid disclosure entirely or set another window, so follow the program's text.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Publishing.&lt;/strong&gt; Wait until the program allows it. Immunefi, for example, prohibits "publicly disclosing a bug report--or even the existence of a bug report for a specific project--before it has been fixed and paid" and requires researchers to follow the project's publication category. When you do publish: describe the flaw and the fix, leave out anything that helps replay it against unpatched users, name no individuals, and never include other people's data.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;If the owner goes silent.&lt;/strong&gt; Follow up through the same channel. Check the policy for an escalation path. Immunefi's rules, for example, refer to requesting mediation. A national CERT can sometimes coordinate if an owner does not respond. Do not escalate to public disclosure before the policy window allows it.&lt;/p&gt;

&lt;h2&gt;
  
  
  8. Checklist
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Find the program on the owner's own page or a platform, and confirm it is current.&lt;/li&gt;
&lt;li&gt;Read the full policy: scope, exclusions, safe harbour, disclosure terms.&lt;/li&gt;
&lt;li&gt;Treat &lt;code&gt;security.txt&lt;/code&gt; as a contact, not a permission.&lt;/li&gt;
&lt;li&gt;Check &lt;code&gt;Expires&lt;/code&gt; and &lt;code&gt;Canonical&lt;/code&gt; in &lt;code&gt;security.txt&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Create your own test accounts and stay inside the rate limits.&lt;/li&gt;
&lt;li&gt;Stop at the minimum proof. Do not keep other people's data.&lt;/li&gt;
&lt;li&gt;Check for known issues and duplicates.&lt;/li&gt;
&lt;li&gt;Write a specific title, exact asset, numbered steps, a working proof of concept and a reasoned impact.&lt;/li&gt;
&lt;li&gt;Confirm every automated finding by hand before submitting.&lt;/li&gt;
&lt;li&gt;Submit through the named channel only.&lt;/li&gt;
&lt;li&gt;Wait for the program's disclosure window before publishing anything.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Sentinelleai posts new bounty programs and reward changes to a free Telegram channel → &lt;a href="https://sentinelleai.app/bounty-alerts?utm_source=devto&amp;amp;utm_medium=organic&amp;amp;utm_campaign=bounty-programs" rel="noopener noreferrer"&gt;https://sentinelleai.app/bounty-alerts?utm_source=devto&amp;amp;utm_medium=organic&amp;amp;utm_campaign=bounty-programs&lt;/a&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>bugbounty</category>
      <category>infosec</category>
      <category>webdev</category>
    </item>
    <item>
      <title>Verified, then swapped: two supply-chain attacks on the code you didn't write</title>
      <dc:creator>Sentinelleai</dc:creator>
      <pubDate>Sat, 07 Nov 2026 18:01:17 +0000</pubDate>
      <link>https://dev.to/sentinelleai/verified-then-swapped-two-supply-chain-attacks-on-the-code-you-didnt-write-3nfh</link>
      <guid>https://dev.to/sentinelleai/verified-then-swapped-two-supply-chain-attacks-on-the-code-you-didnt-write-3nfh</guid>
      <description>&lt;p&gt;On 4 November 2026, two unrelated reports landed on the same day.&lt;/p&gt;

&lt;p&gt;In the morning, Debian, Ubuntu and the developers of APT, the program that installs software on both systems, published fixes for a flaw that let an attacker on the network &lt;strong&gt;swap a package after it had passed its signature check&lt;/strong&gt;. The attacker's code then ran as root.&lt;/p&gt;

&lt;p&gt;A few hours later, the French security firm Sekoia described a campaign it links to the Russian group APT29. For one week, a small &lt;strong&gt;web analytics script loaded by thousands of websites&lt;/strong&gt; quietly added a second script that profiled every visitor's phone and served it an exploit chain to match.&lt;/p&gt;

&lt;p&gt;One story is about servers, the other about websites. They teach the same lesson: most of the code that runs on your machines and in your visitors' browsers was written by someone else, and the checks we rely on cover less than we think.&lt;/p&gt;

&lt;h2&gt;
  
  
  A package, checked and then replaced
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fjo6305lr9fp9qcb09kdt.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fjo6305lr9fp9qcb09kdt.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Debian and Ubuntu do not trust the mirror a package comes from. Every package list is signed with the distribution's key, and every package must match the checksum in that signed list. That is why APT has long been happy to download over plain HTTP: tampering in transit should be caught.&lt;/p&gt;

&lt;p&gt;The flaw, tracked as &lt;strong&gt;CVE-2026-109518&lt;/strong&gt; and reported by researcher Max Justicz, breaks that promise in a narrow but serious way. In the words of the &lt;a href="https://lists.debian.org/deity/2026/11/msg00005.html" rel="noopener noreferrer"&gt;APT developers' advisory&lt;/a&gt;, an attacker can "bypass verification of downloaded packages using redirects and standard transient HTTP server errors that cause APT to retry downloads, leading APT to override existing hardlinks with a chosen payload." &lt;a href="https://lists.debian.org/debian-security-announce/2026/msg00526.html" rel="noopener noreferrer"&gt;Debian's advisory&lt;/a&gt; puts it more simply: improper file handling when resuming a download "allows a man-in-the-middle attacker to replace a package after verification."&lt;/p&gt;

&lt;p&gt;The check happened. The file that was checked is not the file that got installed. And because a .deb package runs its installation scripts as root, the attacker gets the whole machine.&lt;/p&gt;

&lt;p&gt;Three details matter for anyone deciding how worried to be:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Who can do it.&lt;/strong&gt; Anyone who can alter the HTTP traffic between the machine and the package server: a hostile network, a compromised proxy, or a malicious mirror, &lt;a href="https://ubuntu.com/blog/apt-vulnerability-fixes-available" rel="noopener noreferrer"&gt;according to Ubuntu&lt;/a&gt;. It triggers when the machine installs or upgrades a package.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Which systems.&lt;/strong&gt; The attack needs APT to retry failed downloads. That has been the default since APT 2.3.2, which means Debian 12 and later, and Ubuntu 22.04 LTS and later. Older releases are exposed only if someone set &lt;code&gt;Acquire::Retries&lt;/code&gt;, and the APT developers note that "many user scripts" did. Ubuntu rates it High (CVSS 4.0: 7.7) and says derivative distributions are affected too.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What protects you.&lt;/strong&gt; Ubuntu states that HTTPS between the machine and the archive blocks this attack, because the attacker must manipulate plain HTTP traffic. Setting &lt;code&gt;Acquire::Retries&lt;/code&gt; to 0 also prevents it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Fixes are out for every supported branch (Debian 13 "trixie" gets APT 3.0.4; Ubuntu published USN-8996-1, and by 5 November listed every release from 14.04 to 26.10 as fixed). No advisory reports attacks in the wild. Max Justicz has not published a technical write-up yet, so the diagram above is drawn from the advisories' own wording, not from exploit code.&lt;/p&gt;

&lt;h2&gt;
  
  
  A tracker that changed hands for a week
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fruavw9pnvtytyo30ix9o.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fruavw9pnvtytyo30ix9o.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.sekoia.com/blog/poisoned-source-apt29-watering-holes-targets-mobile-devices-with-n-day-exploits" rel="noopener noreferrer"&gt;Sekoia's report&lt;/a&gt;, published the same day, starts with websites the attackers broke into directly: a Ukrainian gas producer, a US chemical industry association, a Bulgarian military-related site, and lookalike copies of sites in the US, Ukraine, the Czech Republic and Bulgaria. Sekoia links the campaign to APT29 with moderate to high confidence.&lt;/p&gt;

&lt;p&gt;Then the investigators noticed requests from sites that had no reason to be targets. The common denominator was the &lt;strong&gt;Plerdy SEO Tracker&lt;/strong&gt;, a third-party analytics script. Between &lt;strong&gt;27 August and 3 September 2026&lt;/strong&gt;, the attackers injected code into Plerdy's own file, &lt;code&gt;a.plerdy.com/public/js/click/main.js&lt;/code&gt;. Every site that loaded it served its visitors two things: the real tracker, and a few lines that created a new &lt;code&gt;&amp;lt;script&amp;gt;&lt;/code&gt; element pointing at a domain the attackers controlled. Sekoia writes that this exposed hundreds of legitimate websites and their visitors, and widened the campaign's reach across thousands of mostly Ukrainian sites.&lt;/p&gt;

&lt;p&gt;What came next was selective. The second script, which the attackers named &lt;em&gt;UltimateFingerprinter&lt;/em&gt;, collected 394 fields about the visitor's device and browser, hidden at the end of a 2.3 MB copy of the AngularJS library. Visitors with the right versions got an exploit chain:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;iOS 16.7&lt;/strong&gt;: two flaws, including CVE-2024-27834, to steal browser cookies;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;iOS 18.5&lt;/strong&gt;: the five-flaw &lt;em&gt;DarkSword&lt;/em&gt; chain, which takes messages, WhatsApp and Telegram databases, the iCloud Keychain, Wi-Fi passwords and photos;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Android, Chrome 139 to 145&lt;/strong&gt; (143 excluded): two flaws that end in a payload Sekoia calls &lt;em&gt;AndroSnitch&lt;/em&gt;, which collects Chrome data and saved Wi-Fi passwords.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Every one of those flaws had been patched before the campaign began. As Sekoia puts it, "a user running an up-to-date operating system is safe from these attacks." On the same 4 November, CISA added &lt;a href="https://www.cisa.gov/known-exploited-vulnerabilities-catalog" rel="noopener noreferrer"&gt;CVE-2024-27834&lt;/a&gt;, fixed by Apple in May 2024, to its catalogue of vulnerabilities known to be exploited, with a deadline of 18 November for US federal agencies.&lt;/p&gt;

&lt;p&gt;The site owners did nothing wrong in the usual sense. They added a popular tool with one line of HTML. That line gave a third party the right to run any code it liked on their pages, for every visitor, and for one week that third party was not who they thought.&lt;/p&gt;

&lt;p&gt;It is not only state groups. In October, Report URI and iVerify described &lt;strong&gt;DeadTag&lt;/strong&gt;: more than 660 expired domains behind forgotten analytics and advertising tags, re-registered by someone else and still embedded on thousands of websites, an estimated 135 million visits a month. Some of that traffic was sold on to the same family of iPhone exploit kits (&lt;a href="https://blog.report-uri.com/deadtag-how-forgotten-website-tags-became-a-delivery-network-for-iphone-spyware/" rel="noopener noreferrer"&gt;Report URI, 22 Oct&lt;/a&gt;).&lt;/p&gt;

&lt;h2&gt;
  
  
  Two months, one week of exposure
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzrfs3akkhr2q5hehnb3j.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzrfs3akkhr2q5hehnb3j.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The gap between the compromise and the public report is typical. Plerdy's file was clean again on 3 September; Sekoia found the campaign in mid-October and published on 4 November. A site owner who waited for the news to check their scripts would have checked two months late. The control has to be in place before anyone knows there is something to look for.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which control would have stopped what
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9p3se6l2lld4tw1vuk6h.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9p3se6l2lld4tw1vuk6h.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Web browsers already have two tools built for this, and most small sites use neither.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Subresource Integrity (SRI)&lt;/strong&gt; pins a script to its exact content. You add a hash to the tag, and the browser refuses to run the file if a single byte changes:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight html"&gt;&lt;code&gt;&lt;span class="nt"&gt;&amp;lt;script &lt;/span&gt;&lt;span class="na"&gt;src=&lt;/span&gt;&lt;span class="s"&gt;"https://cdn.example.com/lib-4.2.1.min.js"&lt;/span&gt;
        &lt;span class="na"&gt;integrity=&lt;/span&gt;&lt;span class="s"&gt;"sha384-…"&lt;/span&gt;
        &lt;span class="na"&gt;crossorigin=&lt;/span&gt;&lt;span class="s"&gt;"anonymous"&lt;/span&gt;&lt;span class="nt"&gt;&amp;gt;&amp;lt;/script&amp;gt;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You can compute the hash yourself: &lt;code&gt;curl -s &amp;lt;url&amp;gt; | openssl dgst -sha384 -binary | openssl base64 -A&lt;/code&gt;. SRI would have refused Plerdy's altered file. The catch is that analytics vendors update their file in place at the same address, so a pinned hash breaks the day they ship a change. For those scripts, the honest options are to host a copy you control and update on purpose, or to accept that you are trusting the vendor completely.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Content-Security-Policy (CSP)&lt;/strong&gt; tells the browser which domains a page may load scripts from. The injected Plerdy code needed to load a second file from a domain the attackers controlled. A policy that only lists the domains you actually use would have blocked that second file (as long as it is a plain list of domains: the &lt;code&gt;'strict-dynamic'&lt;/code&gt; keyword deliberately lets an allowed script load others), and with a &lt;code&gt;report-to&lt;/code&gt; or &lt;code&gt;report-uri&lt;/code&gt; directive it would have told you the attempt happened. CSP does not help if the domain on your list is the one that turns hostile, which is exactly the DeadTag case: there, the fix is to remove tags you no longer use.&lt;/p&gt;

&lt;p&gt;Neither control replaces the oldest one: &lt;strong&gt;know what your pages load&lt;/strong&gt;. Most sites have never listed their third-party scripts, so nobody notices when one of them changes or its domain expires.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you can check tonight
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftwf3l7of2ue5460qujwo.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftwf3l7of2ue5460qujwo.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;If you run Debian or Ubuntu servers&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Check your APT version&lt;/strong&gt; against the fixed versions in &lt;a href="https://ubuntu.com/blog/apt-vulnerability-fixes-available" rel="noopener noreferrer"&gt;Ubuntu's table&lt;/a&gt; or the &lt;a href="https://security-tracker.debian.org/tracker/CVE-2026-109518" rel="noopener noreferrer"&gt;Debian tracker&lt;/a&gt;: &lt;code&gt;apt list --installed '?source-package(^apt$)'&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Upgrade with retries switched off&lt;/strong&gt;, as Ubuntu recommends, so the update itself cannot be tampered with: &lt;code&gt;sudo apt update &amp;amp;&amp;amp; sudo apt upgrade -o Acquire::Retries=0&lt;/code&gt;. Unattended upgrades, on by default on Ubuntu, will apply it within a day if you leave them on.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Move your package sources to HTTPS.&lt;/strong&gt; Replace &lt;code&gt;http://&lt;/code&gt; with &lt;code&gt;https://&lt;/code&gt; in &lt;code&gt;/etc/apt/sources.list&lt;/code&gt; and &lt;code&gt;/etc/apt/sources.list.d/&lt;/code&gt;, then run &lt;code&gt;sudo apt update&lt;/code&gt; and check for errors: some community mirrors do not support HTTPS.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Do not forget the machines you never log into.&lt;/strong&gt; Container images, CI runners and build servers run APT too. An image built before 4 November carries the old APT; rebuild it from an updated base.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;If you run websites, yours or your clients'&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;List every third-party script&lt;/strong&gt; each site loads: open the browser's developer tools, Network tab, filter on JS, and reload. Or from a terminal: &lt;code&gt;curl -s https://your-site | grep -oE '&amp;lt;script[^&amp;gt;]+src="[^"]+"'&lt;/code&gt; (this misses scripts added later by other scripts, so the browser view is the one to trust).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;For each one, ask three questions.&lt;/strong&gt; Is it still used? Is its domain still registered by the vendor you think? Could you host a fixed copy yourself?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Add SRI to every script with a fixed version&lt;/strong&gt;, from a CDN or anywhere else. Tools such as srihash.org generate the tag.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Write a Content-Security-Policy&lt;/strong&gt; that lists only the script domains you use, start it in report-only mode to see what breaks, and send reports somewhere a person reads them. Google's CSP Evaluator and Mozilla's HTTP Observatory are free ways to grade the result.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Update phones and browsers.&lt;/strong&gt; All three exploit chains in this campaign used flaws that were already fixed.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The lesson
&lt;/h2&gt;

&lt;p&gt;Signatures, checksums and well-known vendors are real protections, and they all share one assumption: that the thing you checked is the thing that runs. This week showed two ways that assumption fails, one on a server and one in a browser. Neither fix is exotic. Turn on HTTPS, update the package manager, list your scripts, pin what can be pinned, and tell the browser where code is allowed to come from.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;I build &lt;a href="https://sentinelleai.app/for/agencies?utm_source=hashnode&amp;amp;utm_medium=article&amp;amp;utm_campaign=article-code-you-didnt-write" rel="noopener noreferrer"&gt;Sentinelleai&lt;/a&gt;, which reads the repositories you deliver for vulnerable and malicious dependencies and checks the CDN scripts your pages load for subresource integrity.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Sources: Debian Security Advisory DSA-6607-1 and APT-2026-11-01 security update (4 Nov 2026); Ubuntu, "CVE-2026-109518 APT vulnerability fixes available" and USN-8996-1; Sekoia TDR, "Poisoned Source: APT29 watering holes target mobile devices with n-day exploits" (4 Nov 2026); CISA Known Exploited Vulnerabilities catalogue (4 Nov 2026); Report URI and iVerify, "DeadTag" (22 Oct 2026). Details may change as researchers publish more.&lt;/em&gt;&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fga4du8c22qt8t50byev0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fga4du8c22qt8t50byev0.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>linux</category>
      <category>webdev</category>
      <category>cybersecurity</category>
    </item>
    <item>
      <title>A laptop, a container and one signature: the $6.3B Binance drain, in three numbers</title>
      <dc:creator>Sentinelleai</dc:creator>
      <pubDate>Mon, 02 Nov 2026 18:02:50 +0000</pubDate>
      <link>https://dev.to/sentinelleai/a-laptop-a-container-and-one-signature-the-63b-binance-drain-in-three-numbers-34f7</link>
      <guid>https://dev.to/sentinelleai/a-laptop-a-container-and-one-signature-the-63b-binance-drain-in-three-numbers-34f7</guid>
      <description>&lt;p&gt;No smart contract was broken. What happened, what it cost, and what anyone who signs transactions can check tonight.&lt;/p&gt;

&lt;p&gt;On 29 October 2026, starting at 20:05 UTC, Binance wallets on ten blockchains sent their balances to addresses controlled by a single attacker. Most of it left in about twenty minutes. It is the largest crypto theft on record, roughly four times the $1.5 billion Bybit hack of February 2025.&lt;/p&gt;

&lt;p&gt;No smart contract was exploited. Every outflow carried a valid Binance signature. That is the part worth understanding, because it applies far beyond one exchange.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three numbers, and why they don't add up neatly
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fny9na4b5b152sez5y0bz.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fny9na4b5b152sez5y0bz.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;About $6.3 billion taken.&lt;/strong&gt; That is Binance's own estimate of the total impact (&lt;a href="https://www.binance.com/en/support/announcement/detail/c5a8729a97ac486d963784d96c5642d7" rel="noopener noreferrer"&gt;Security Incident Update, 30 Oct&lt;/a&gt;). TRM Labs, valuing each transfer at the price of the moment, counts &lt;a href="https://www.trmlabs.com/resources/blog/north-korea-steals-usd-6-36-billion-from-binance-in-the-largest-hack-in-history" rel="noopener noreferrer"&gt;$6.36 billion across ten chains&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;About $3 billion frozen.&lt;/strong&gt; Stablecoins have issuers, and issuers can freeze. Binance puts the frozen total at about $3 billion. TRM's on-chain count at 00:30 UTC on 30 October was about $2.7 billion: roughly 1.77 billion USDT frozen by Tether, 825 million USDC by Circle, 40.5 million RLUSD by Ripple, 27.8 million USD1 by World Liberty Financial, and about $30 million in ETH and ARB moved out of the attacker's reach by Arbitrum's Security Council. That leaves a little over $3 billion that the freezes did not stop.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;About $2.4 billion beyond anyone's reach.&lt;/strong&gt; Bitcoin and ether have no issuer. Racing the freezes, the attacker sold about 325 million USDT for ether at a discount of more than 10% and borrowed ether against USDT on a lending protocol. TRM's count: more than 420,000 ETH and 18,270 BTC, about $2.4 billion, that no freeze can touch.&lt;/p&gt;

&lt;p&gt;The figures come from different sources, at different hours, with different price references, and the attacker lost value on every hurried conversion. They will not add up to the dollar, and anyone who tells you they do is rounding.&lt;/p&gt;

&lt;p&gt;There is a fourth number, and it is the one customers care about. Binance says it will absorb the loss with its emergency fund (SAFU), part of its futures insurance fund and its own treasury. On 31 October it reported &lt;a href="https://www.binance.com/en/support/announcement/detail/1fdf1f61559c401da0435895ffcccb62" rel="noopener noreferrer"&gt;90% of the stolen assets replenished&lt;/a&gt;; on 2 November, &lt;a href="https://www.binance.com/en/support/announcement/detail/185bd9a4717f44088b3be70287b8237f" rel="noopener noreferrer"&gt;95%&lt;/a&gt;. The remainder is less liquid tokens that take longer to buy. "Replenished" is not "recovered": the exchange paid with its own money, and the stolen coins are still where the attacker put them.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it happened
&lt;/h2&gt;

&lt;p&gt;Binance's initial investigation says the incident "may involve a compromised DevOps engineer's laptop, which allowed an attacker to gain access to a service container and submit unauthorized withdrawal transactions." It found no evidence that a wallet private key leaked. A full post-mortem is still to come.&lt;/p&gt;

&lt;p&gt;That fits what the chain shows. Wallets on ten chains, with different key types and signing schemes, all produced valid signatures for the attacker within about half an hour. One leaked key does not do that. Access to the system that asks for signatures does.&lt;/p&gt;

&lt;p&gt;On Ethereum, the attacker added a step. It used &lt;strong&gt;EIP-7702&lt;/strong&gt;, a feature introduced with Ethereum's Pectra upgrade in May 2025.&lt;/p&gt;

&lt;h2&gt;
  
  
  EIP-7702 in two paragraphs
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F207b1o72pcewwqzj4imk.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F207b1o72pcewwqzj4imk.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;An ordinary Ethereum address (an "externally owned account") has no code; it just signs. EIP-7702 lets that address sign an &lt;em&gt;authorization&lt;/em&gt; that points it at a smart contract. From then on, any call to the address runs that contract's code, with the address's full balance and permissions. The arrangement stays until the address signs a new authorization.&lt;/p&gt;

&lt;p&gt;That is useful: it is how wallets add batching, spending limits or gas sponsorship without moving funds. It also means that &lt;strong&gt;one signature can hand an address to someone else, permanently, until it is revoked&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Ethereum takeover, minute by minute
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9zaj0pwyjb0jedgvzypb.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9zaj0pwyjb0jedgvzypb.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The preparation was visible days ahead. The attacker's operator wallet was funded through Tornado Cash about 36 hours before. On 29 October it deployed a "sweeper" contract, tested it on throwaway wallets, rehearsed, and deployed the final version at 15:59 UTC. That contract accepts instructions only from the attacker, and can move every token out of whatever address runs it.&lt;/p&gt;

&lt;p&gt;At 20:20:59 UTC, one transaction carried &lt;strong&gt;50 authorizations&lt;/strong&gt;, each signed by Binance's main Ethereum hot wallet (the address labelled "Binance 14"), each pointing it at the sweeper. Why 50? An authorization is valid only at the wallet's current transaction count (its nonce), and a busy exchange wallet's count moves every few seconds. Signing 50 consecutive counts in advance guaranteed one would match. One did.&lt;/p&gt;

&lt;p&gt;Three minutes and 24 seconds later, a single call moved &lt;strong&gt;167,464 ETH and 63 tokens&lt;/strong&gt;, including about 1.6 billion USDT.&lt;/p&gt;

&lt;p&gt;Two details make this attack instructive rather than merely large:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The delegation stayed in place.&lt;/strong&gt; Customer deposits kept arriving, and the attacker swept the wallet four more times, until Binance revoked the delegation at 00:10 UTC on 30 October. As &lt;a href="https://www.hypernative.io/insights/blog/inside-the-largest-crypto-theft-of-all-time-the-6-3b-binance-hot-wallet-drain" rel="noopener noreferrer"&gt;Hypernative notes&lt;/a&gt;, the delegation belongs to the address, not to the key: rotating the key would not have removed it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The delegation inherited everything the wallet owned.&lt;/strong&gt; The hot wallet was the sole owner of a separate smart account holding USDC. Once the attacker controlled the wallet, it controlled that account too, and pulled about $823 million USDC out of it. Circle froze it shortly afterwards.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  "But what if the laptop story is wrong?"
&lt;/h2&gt;

&lt;p&gt;It might be incomplete. Binance's own wording is "may involve", and the post-mortem is not out yet. It does not matter much for the lesson. Every element of this scenario is ordinary:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a workstation belonging to someone with production access;&lt;/li&gt;
&lt;li&gt;a service that can request signatures without a human looking at each one;&lt;/li&gt;
&lt;li&gt;a signature format, EIP-7702, that hands over an address in one step;&lt;/li&gt;
&lt;li&gt;no automatic limit on how much can leave, or to where.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Each of these exists in thousands of companies and teams far smaller than Binance. Whether this particular chain of events is confirmed or revised, the chain itself is possible, and that is reason enough to check your own.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you can check tonight
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpninim2pu8u2ft3jwivd.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpninim2pu8u2ft3jwivd.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;If your team signs from a hot wallet or a treasury wallet&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Look for code at your addresses.&lt;/strong&gt; An address that should be a plain wallet should have no code. With Foundry installed: &lt;code&gt;cast code &amp;lt;address&amp;gt; --rpc-url &amp;lt;your RPC&amp;gt;&lt;/code&gt;. An empty answer (&lt;code&gt;0x&lt;/code&gt;) is normal. An answer starting with &lt;code&gt;0xef0100&lt;/code&gt; means the address is delegated, and the 20 bytes that follow are the contract it now runs. Put that check on an alert, not on a to-do list.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Refuse EIP-7702 authorizations by policy&lt;/strong&gt; unless you expect one. A request for many authorizations across a range of nonces is not a configuration quirk; TRM calls it an indicator of compromise.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Know how to revoke before you need to.&lt;/strong&gt; An authorization that points to the zero address clears a delegation. Prepare it, test it on a test network, and know who signs it at 3 a.m.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;List what your hot wallet owns.&lt;/strong&gt; Smart accounts, admin roles, allowances: a delegation inherits all of them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Put limits on outflows.&lt;/strong&gt; A transfer of an entire balance to an address that did not exist yesterday should be stopped automatically, not reviewed the next morning. That is what circuit breakers, withdrawal velocity limits and pause roles on contracts are for.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Separate signing paths&lt;/strong&gt; by chain and by wallet tier, so one compromise cannot reach ten chains in half an hour.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;If you run any company where someone's laptop can reach production&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fa8ec339dfbfk6mk10129.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fa8ec339dfbfk6mk10129.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The laptop is the front door. The usual controls exist and are not exotic: endpoint detection and response on every admin machine (commercial suites such as Microsoft Defender for Endpoint, CrowdStrike Falcon or SentinelOne, or open-source monitoring such as Wazuh and osquery); hardware security keys (FIDO2) instead of codes that can be phished; a dedicated machine for administrative access, kept separate from email and browsing; and short-lived, just-in-time access to production rather than standing credentials, as tools like Teleport or HashiCorp Boundary provide.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;If you simply hold crypto in your own wallet&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;EIP-7702 applies to you too. A phishing page no longer needs your seed phrase; it can ask you to sign an "account upgrade". Read every signature request, and be wary of anything that mentions delegation, smart account upgrades or authorizations you did not start. An authorization signed with chain ID 0 is valid on every EVM chain at once. You can check whether your own address is delegated with a block explorer or the &lt;code&gt;cast code&lt;/code&gt; command above.&lt;/p&gt;

&lt;h2&gt;
  
  
  The lesson
&lt;/h2&gt;

&lt;p&gt;Binance survived this because it could pay $6 billion out of its own pocket and because stablecoin issuers froze about half of it within an hour. Almost nobody else can count on either. For everyone else, the protections that matter are the boring ones: knowing when your own addresses change, limiting what can leave and how fast, and treating the laptop of anyone with signing access as part of the vault.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;I build &lt;a href="https://sentinelleai.app/for/web3?utm_source=medium&amp;amp;utm_medium=article&amp;amp;utm_campaign=article-binance-eip7702" rel="noopener noreferrer"&gt;Sentinelleai&lt;/a&gt;, which watches the servers you enroll for signs of compromise and can pause a contract you own when an attack starts.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Sources: Binance Security Incident Update (30 Oct) and Recovery Updates #3 (31 Oct) and #5 (2 Nov); TRM Labs (29 Oct, figures as of 30 Oct 00:30 UTC); Hypernative (30 Oct); EIP-7702 specification (eips.ethereum.org/EIPS/eip-7702). Figures may change as investigations continue.&lt;/em&gt;&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9v2lalbtw2ihlcngpof4.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9v2lalbtw2ihlcngpof4.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://medium.com/@bosacks_25667/a-laptop-a-container-and-one-signature-the-6-3b-binance-drain-in-three-numbers-31f682e8b2c3?sharedUserId=bosacks_25667" rel="noopener noreferrer"&gt;https://medium.com/@bosacks_25667/a-laptop-a-container-and-one-signature-the-6-3b-binance-drain-in-three-numbers-31f682e8b2c3?sharedUserId=bosacks_25667&lt;/a&gt;&lt;/p&gt;

</description>
      <category>cybersecurity</category>
      <category>ethereum</category>
      <category>cryptocurrency</category>
      <category>blockchain</category>
    </item>
    <item>
      <title>One second of block time: replaying the Fence Finance TranchePool drain</title>
      <dc:creator>Sentinelleai</dc:creator>
      <pubDate>Wed, 28 Oct 2026 15:52:31 +0000</pubDate>
      <link>https://dev.to/sentinelleai/one-second-of-block-time-replaying-the-fence-finance-tranchepool-drain-hni</link>
      <guid>https://dev.to/sentinelleai/one-second-of-block-time-replaying-the-fence-finance-tranchepool-drain-hni</guid>
      <description>&lt;p&gt;On 27–28 October 2026, a TranchePool on Avalanche associated with Fence Finance, Valinor Digital and Oatfi lost 1,013,353 USDC, its entire balance. Within hours two readings circulated. Blockaid described flash-loan minting of inflated tranche shares. SlowMist described a settlement bug at the UTC day rollover. We wanted to know which one the chain agrees with, so we replayed the attack on a local fork.&lt;/p&gt;

&lt;h2&gt;
  
  
  The two transactions
&lt;/h2&gt;

&lt;p&gt;At 23:56:41 UTC the attacker's contract took a flash loan of 1,013,353.61 USDC from Aave v3. That is exactly the pool's balance. Inside the callback it made five calls to the pool. One of them carries the same amount and a timestamp, &lt;code&gt;0x6ae13b00&lt;/code&gt;, which is 00:00:00 UTC on 28 October. The last call draws the full amount back out. Aave is repaid with a 506 USDC premium. After this transaction the pool still holds its 1,013,353 USDC. Nothing has been stolen yet.&lt;/p&gt;

&lt;p&gt;At 00:00:03 UTC, three seconds after midnight, the attacker called the same draw again. This time the pool paid out its real USDC.&lt;/p&gt;

&lt;h2&gt;
  
  
  The fork replay
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F0i39ftluqlpj0abirkgv.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F0i39ftluqlpj0abirkgv.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;We forked Avalanche C-Chain at the block just before midnight and impersonated the attacker. Then we sent the second transaction's exact calldata twice, changing only the block time:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;at 23:59:59 the draw reverts;&lt;/li&gt;
&lt;li&gt;at 00:00:00 it succeeds and the pool goes from 1,013,353.61 USDC to zero.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A fork at the real block, 00:00:03, reproduces the on-chain result to the unit.&lt;/p&gt;

&lt;p&gt;The same state and the same calldata were used in every run. The only variable is the clock, so the draw check depends on time. That fits SlowMist's description: at the rollover, the pool's settlement logic treated the scheduled principal payment as received and subtracted it from outstanding principal, although no USDC had arrived. The draw check, funded principal minus outstanding principal, then reported a full line of capacity again.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who was right
&lt;/h2&gt;

&lt;p&gt;Both readings were partly right. A flash loan was used, but only to set the position up without putting capital at risk. The theft itself is the rollover: a due date was booked as a payment.&lt;/p&gt;

&lt;h2&gt;
  
  
  The lesson for anyone building credit or vault logic
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpdyqru2ooegxnr2pcnrg.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpdyqru2ooegxnr2pcnrg.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Real-world credit brings real-world conventions on-chain: due dates, day counts, waterfalls. Each of them is a place where state can move without money moving. The invariant that holds the line is short:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Draw capacity can never exceed the assets the contract has actually received.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Write it down, then fuzz it across the boundaries where time changes the answer:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;the day rollover;&lt;/li&gt;
&lt;li&gt;a due date equal to the current block timestamp;&lt;/li&gt;
&lt;li&gt;a due date chosen by the caller.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Status
&lt;/h2&gt;

&lt;p&gt;Fence has sent the attacker an on-chain message offering a 10% recovery bounty for the return of 90% of the funds.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Addresses and transactions:&lt;/em&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;pool &lt;code&gt;0xe25cb545bdd47a8ec2d08001cb5661b00d47621a&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;attacker &lt;code&gt;0x6ff1c3ffa622e1062168f3497a275be9cc2b1246&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;setup tx &lt;code&gt;0x19ab145e7f771549ca2f1da4b1768d7f728834015b2480ddb5be19c9769899fd&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;drain tx &lt;code&gt;0x3f6b23292a2e9c2956c7ac56d0b1f79b0271b49edec2156d2675ddf1de058c77&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The pool's source is not verified on the explorers, so this analysis rests on calldata, events, balances and the fork replay.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Sentinelleai reads smart contracts before launch, fuzzes declared invariants, and watches deployed contracts after. &lt;a href="https://sentinelleai.app/for/web3?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=fence" rel="noopener noreferrer"&gt;See how it works for Web3 teams.&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>web3</category>
      <category>blockchain</category>
      <category>solidity</category>
    </item>
  </channel>
</rss>
