Cisco's email gateway was exploited via malicious SQL embedded in email. Mandiant confirmed attackers now weaponise vulnerabilities in roughly 5 days while the median org takes 43 to patch. A China-linked group chained Chrome and Windows zero-days to deploy a new backdoor. And an AI coding assistant was hijacked to spread malware across 100 internal repos."
- Cisco Secure Email Gateway zero-day (CVSS 9.8, CVE-2026-76461) — attackers send a crafted email containing malicious SQL, get root on the appliance, no credentials required
- China-linked UTA0560 chained Chrome and Windows zero-days to deliver GRIMWEDGE, a JavaScript backdoor, against NGOs in a targeted spear-phishing campaign
- WSO2 API Manager JWT bypass (CVSS 9.8, CVE-2026-5430) under active exploitation — attackers are forging admin tokens and taking over accounts
- Mandiant AI Threat Tracker: attackers now weaponise new vulnerabilities in about 5 days; the median organisation takes 43 days to patch
- An AI coding assistant session was hijacked at a SaaS provider, spreading Shai-Hulud malware across about 100 internal code repositories
- CrowdSec breach — 170 private GitHub repos copied using credentials stolen from a departed employee whose access was never revoked
- NightmareStresser DDoS-for-hire taken down — one of the longest-running DDoS services ever, active since at least 2022
- US/UK/Netherlands joint advisory on Iranian intelligence Windows malware targeting dissidents, journalists, and activists
Cisco Secure Email Gateway: root access via a malicious email
CVE-2026-76461 was exploited before Cisco disclosed and patched it, and allows unauthenticated, remote attackers to execute commands with root privileges on vulnerable systems. The mechanism is about as straightforward as it gets: attackers embed malicious SQL statements directly into inbound email payloads, and when the appliance's AsyncOS email parsing logic processes the crafted message, the queries execute unchecked, escalating directly to full root operating system access.
CISA added CVE-2026-76461 to its KEV catalog on September 14, ordering federal agencies to apply patches by September 17. The same patch batch from Cisco addressed four other critical flaws in Secure Email Gateway and Secure Email and Web Manager, though those are not yet confirmed exploited. Cloud-hosted instances were remediated server-side by Cisco; if you run the appliance on-premises, this one is on you.
Cisco also published indicators of compromise and recommended monitoring mail logs for suspicious SQL statements and cross-checking network logs for unusual external traffic. Given that the exploit literally arrives via email, any internet-facing SEG that hasn't been patched this week should be treated as potentially compromised.
GRIMWEDGE: China-linked UTA0560 chains Chrome and Windows zero-days
A Chinese threat actor tracked as UTA0560 ran a spear-phishing campaign exploiting recently patched security flaws in Google Chrome and Microsoft Windows to deliver GRIMWEDGE, a malicious JavaScript backdoor. The activity targeted multiple NGOs. The chain being exploited links last week's patched Chrome V8 zero-day with one of the Windows elevation-of-privilege bugs from September's Patch Tuesday — a reminder that "no exploitation observed at time of patch" is a time-limited status, not a permanent one. The broader BlueMoon exploit kit is facilitating wider adoption of the same chain by multiple espionage-motivated groups.
WSO2 API Manager JWT bypass under active exploitation
CVE-2026-5430 (CVSS 9.8) in WSO2 API Manager is a case of improper verification of a cryptographic signature. JWT authentication can be bypassed when a token is signed using an unsupported algorithm, allowing unauthorised access — and watchTowr's honeypot captured JWT tokens arriving on September 13 with baked-in administrator privileges. Successful exploitation means full account takeover and access to whatever that API gateway protects downstream. WSO2 disclosed the flaw in May 2026; active exploitation began three months later. If WSO2 API Manager sits anywhere near your AI infrastructure, authentication layer, or data pipelines, treat this as urgent.
Mandiant: 5 days to weaponise, 43 days to patch
Attackers now weaponise new vulnerabilities in about five days. The median organisation takes 43 days to patch one. Those two numbers, from Mandiant's AI Threat Tracker and Verizon's DBIR 2026 respectively, define the problem more precisely than almost anything else published this week. The eight-to-one gap between how fast exploits develop and how fast patches get applied is the operational reality defenders are working inside.
The same report confirmed what Oligo Security attributed in August: Mandiant formally identified TeamPCP as UNC6780 and linked the group's activity since March 2026 to large-scale supply-chain compromises across PyPI, npm, and Docker Hub. The Mandiant tracker also documented threat actors using LLMs for hyper-personalised social engineering, malware that queries AI models mid-execution to evade detection, and Q2 2026 data theft operations where attackers specifically targeted proprietary AI models, training data, and research, including from healthcare organisations.
AI coding assistant hijacked, Shai-Hulud spread across 100 repos
An attacker hijacked an active AI coding-assistant session at an unnamed SaaS provider and later spread Shai-Hulud malware across about 100 internal code repositories. Before the repository spread, the assistant had been used for a routine coding task — the attacker laterally pivoted from the compromised session to the connected codebase. Shai-Hulud is the same malware family linked to TeamPCP's npm worm campaign from early August, which makes this the third time in six weeks we have documented it appearing in a new context. This one is worth flagging to any developer team using AI coding assistants with write access to internal repositories: the assistant's permissions are now part of your blast radius.
CrowdSec breach — 170 repos, offboarding failure
An attacker copied about 170 of CrowdSec's private GitHub repositories on May 22 using the account of an employee who had just left; CrowdSec had kept his GitHub access open. CrowdSec says his laptop was compromised in May's supply-chain attack on TanStack, in which malicious npm packages stole credentials from developers' machines. The code appeared on an online forum on September 16. CrowdSec says no infrastructure or databases were accessed and no code was modified — but 83 user email addresses and investment data from 51 potential investors were included in the exfiltrated repos.
The root cause is neither exotic nor new: a departed employee's access token was never revoked. This is the same access-hygiene failure that produced the Okta breach in 2022 and multiple others since. For a security company that makes threat intelligence tooling, this carries a particular sting.
NightmareStresser DDoS-for-hire service taken down
NightmareStresser was one of the longest-running DDoS-for-hire services in the world, active since at least 2022. The takedown comes via coordinated law enforcement action; domain seizures and arrests were confirmed on September 18. DDoS-for-hire services remain a persistent infrastructure problem because the bar to entry for launching an attack is subscription pricing and a target's IP address. Removing a multi-year platform is a meaningful disruption even if the ecosystem around it persists.
Also worth a skim
- Iran's Windows malware — a joint advisory from US, UK, and Netherlands cybersecurity agencies detailed a Windows-based spyware tool used by Iranian intelligence to target dissidents, journalists, and activists globally
- BambooToken malware uses MQTT as its command-and-control channel to control Windows and Linux systems, targeting organisations in Asia and South America — active since at least 2023, only now coming to wider attention
- WooCommerce Wholesale Lead Capture flaw being exploited to plant PHP web shells on WordPress sites; over 6,000 active installs affected
- Vite dev servers being abused to harvest AWS and Azure credentials left exposed in frontend build environments
- SolarWinds Access Rights Manager reported to be affected by an unauthenticated RCE vulnerability
- Arctic Wolf documented large-scale Fortinet VPN credential-stuffing: tens of millions of authentication failures across US environments in late August, two sustained waves over three days
- Marimo RCE — a human attacker exploiting the vulnerability reached an SSH bastion in eight seconds, used as a benchmark to illustrate the speed gap that AI-assisted exploitation is narrowing further
- Anthropic published analysis stating AI misuse is entering a new phase: shifting from cybercrime toward surveillance, propaganda, and weapons-related applications
- OpenAI Codex sandbox — researchers demonstrated a method to escape the Codex sandbox and execute commands on the underlying host
Closing thought
The Mandiant stat is the one to carry into every security budget conversation this quarter: 5 days to weaponise, 43 days to patch. That gap is structural, and the PaperCut campaign from last week put a real number on what it costs -- 11 organisations compromised in 26 seconds. This week's Cisco SEG zero-day, the WSO2 JWT bypass, and the GitLab exploitation from last week all ran the same script: vulnerability disclosed or patched, exploitation began days or hours later, organisations scrambling to catch up.
The AI coding assistant hijack is the other story to sit with. It is not a sophisticated zero-day. It is an attacker getting into a developer's session and using the assistant's write access to internal repos as an amplifier. The tool did exactly what it was given permission to do -- and that's the point.
Sources: BleepingComputer, The Hacker News, CyberScoop, SecurityWeek, Help Net Security, Cybersecurity Dive, WIU Cybersecurity Center, Mandiant/Google Cloud, watchTowr, Privacy Needle, Security Affairs, Hendryadrian.com.
Top comments (0)