<?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: Tiến Thành bình</title>
    <description>The latest articles on DEV Community by Tiến Thành bình (@tin_thnhbnh_6d6c9102d).</description>
    <link>https://dev.to/tin_thnhbnh_6d6c9102d</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%2F4212637%2Ff7cb1dc3-dbaf-410d-abda-7c678a915490.png</url>
      <title>DEV Community: Tiến Thành bình</title>
      <link>https://dev.to/tin_thnhbnh_6d6c9102d</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/tin_thnhbnh_6d6c9102d"/>
    <language>en</language>
    <item>
      <title>Anthropic Goes Public This Week: What the $70–$78 IPO, the ANT Ticker, and a $26B Side Deal Mean for Claude Builders</title>
      <dc:creator>Tiến Thành bình</dc:creator>
      <pubDate>Wed, 18 Nov 2026 02:20:53 +0000</pubDate>
      <link>https://dev.to/tin_thnhbnh_6d6c9102d/anthropic-goes-public-this-week-what-the-70-78-ipo-the-ant-ticker-and-a-26b-side-deal-mean-39jh</link>
      <guid>https://dev.to/tin_thnhbnh_6d6c9102d/anthropic-goes-public-this-week-what-the-70-78-ipo-the-ant-ticker-and-a-26b-side-deal-mean-39jh</guid>
      <description>&lt;p&gt;The company behind Claude is about to become a public company. Anthropic filed its third amended S-1 on November 17, Nasdaq has approved the listing, and the deal is expected to price today, November 18, with trading to begin November 19 — those two dates are reported by CNBC, citing people familiar with the deal; the filing itself sets no date.&lt;/p&gt;

&lt;p&gt;If you build on Claude, this is not just finance news. The filing is the first time Anthropic has put audited numbers, its governance, and its compute bill on the public record. Here is what it actually says — and what it means for developers.&lt;/p&gt;

&lt;h2&gt;
  
  
  The deal in numbers
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;825,000,000 Class A shares&lt;/strong&gt; — 801,351,354 sold by Anthropic, 23,648,646 by existing stockholders&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;$70.00 to $78.00 per share&lt;/strong&gt;, set in the November 10 amendment&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Nasdaq Global Select Market, ticker ANT&lt;/strong&gt; — approved to list, per the November 17 amendment&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Up to $64.35 billion&lt;/strong&gt; in gross proceeds at the top of the range ($57.75 billion at the bottom); the underwriters hold an option on 123,750,000 more shares, which would take the total to about $74 billion&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lead underwriters:&lt;/strong&gt; Goldman Sachs and Morgan Stanley&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Implied value:&lt;/strong&gt; on the filing's count of about 13.09 billion shares outstanding after the offering, the range implies roughly $916 billion at $70 and $1.02 trillion at $78&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That would make it the largest IPO on record by proceeds, more than double the previous biggest.&lt;/p&gt;

&lt;h2&gt;
  
  
  The strange governance — and why builders should care
&lt;/h2&gt;

&lt;p&gt;The numbers are the least interesting part of the filing. The governance is.&lt;/p&gt;

&lt;p&gt;Anthropic is a &lt;strong&gt;Delaware public benefit corporation&lt;/strong&gt;. By law it must balance shareholder returns, the interests of those materially affected by its conduct, and its stated public benefit purpose: "the responsible development and maintenance of advanced AI for the long-term benefit of humanity."&lt;/p&gt;

&lt;p&gt;Three mechanisms are designed to keep that mission alive after the IPO:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The Long-Term Benefit Trust (LTBT).&lt;/strong&gt; An independent body of trustees with expertise in AI safety, national security, public policy, and social enterprise. Through a special class of stock, it elects and removes the directors who will form a &lt;strong&gt;majority of the board&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Class F stock.&lt;/strong&gt; A single class held by a company owned by the seven co-founders carries &lt;strong&gt;50.1% of the total voting power&lt;/strong&gt; on matters submitted to stockholders generally — but no economic rights. The founders keep control of the company; they don't get extra money for it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Class E stock for employees&lt;/strong&gt;, with voting rights only on limited matters, and &lt;strong&gt;Class N stock&lt;/strong&gt; — non-voting shares held by strategic partners, which convert to Class A if sold to an outside party.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;In short: public shareholders get the economics, but not control. For anyone whose product depends on Claude, that structure is the whole point — it is designed so that a quarter of bad earnings can't force a change of direction. Whether it works in practice is now a live experiment, and it will be watched in public.&lt;/p&gt;

&lt;h2&gt;
  
  
  The $26 billion side deal: the compute supply chain buys in
&lt;/h2&gt;

&lt;p&gt;Alongside the IPO, seven companies agreed to buy &lt;strong&gt;$26.0 billion of Class N non-voting shares at the IPO price&lt;/strong&gt; in a concurrent private placement: AMD, Broadcom, Google, Micron, Microsoft, NVIDIA, and TSMC Partners.&lt;/p&gt;

&lt;p&gt;Read that list again. It is, almost exactly, the list of companies Anthropic buys its compute from. Its suppliers are now also its shareholders — and after the deal, Amazon will own about 14.7% of Anthropic's capital stock and Google about 13.7%, per the November 17 amendment.&lt;/p&gt;

&lt;p&gt;That is a bet in both directions: the chip and cloud giants are underwriting Anthropic's demand, and Anthropic is locking in the supply it needs to keep growing.&lt;/p&gt;

&lt;h2&gt;
  
  
  The money picture: $81 billion run-rate vs. a $901 billion compute bill
&lt;/h2&gt;

&lt;p&gt;The filing's financials explain why the raise is so large:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Run-rate revenue of about $81 billion&lt;/strong&gt; as of October 31, 2026, up $10 billion from September 30. Anthropic defines it as trailing 28-day consumption fees multiplied by 13, plus monthly subscription fees multiplied by 12 — and warns it is not GAAP revenue and not a forecast.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GAAP revenue of $33.24 billion&lt;/strong&gt; in the first nine months of 2026, against $4.585 billion in all of 2025. Third-quarter revenue alone was about $17 billion.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;About $900.9 billion in future payment commitments&lt;/strong&gt;, substantially all of it for compute and infrastructure — roughly 11 times the run-rate. About $70.7 billion of it falls due in the 12 months after September 30.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So the IPO is not a victory lap. It is the funding round for a compute bill that dwarfs current revenue. The ~$85 billion of net proceeds from the IPO and the private placement would lift cash and short-term investments to roughly $216 billion on a pro forma basis — still less than a quarter of the commitments.&lt;/p&gt;

&lt;h2&gt;
  
  
  What changes for developers — and what doesn't
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Nothing changes in the API tomorrow.&lt;/strong&gt; Pricing, rate limits, model availability, and the terms of service are not set by the S-1.&lt;/p&gt;

&lt;p&gt;What does change:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Quarterly disclosure.&lt;/strong&gt; Anthropic's revenue, margins, compute spend, and customer concentration will now be public every quarter. The filing already discloses that more than 11,900 customers pay $100,000 or more a year, and that those customers produced about 74% of revenue in the first nine months of 2026.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Public scrutiny of the safety mission.&lt;/strong&gt; The LTBT and the public benefit charter will be tested in the open. Every future model release will be read against the governance structure Anthropic just filed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A bigger war chest — and a bigger bill.&lt;/strong&gt; The $26 billion private placement plus the IPO proceeds fund the compute commitments. For builders, that means capacity expansion is funded, but also that Anthropic's growth now has to keep pace with $70 billion a year of payments coming due.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  One warning worth repeating
&lt;/h2&gt;

&lt;p&gt;Anthropic has said that any third party selling its stock to the public — including through tokens — is likely offering something with no value, and that unapproved sales are void. The shares don't trade yet, and there is no official pre-IPO market. If someone offers you "ANT" before November 19, it's a scam.&lt;/p&gt;

&lt;h2&gt;
  
  
  The bottom line
&lt;/h2&gt;

&lt;p&gt;Anthropic is going public with a $70–$78 range, a ticker, and a governance structure no other frontier lab has tried at this scale: public money, founder control, and a trust with a board majority whose job is to protect the mission, not the quarter.&lt;/p&gt;

&lt;p&gt;For developers, the practical takeaway is simple. Your API provider is about to file quarterly earnings. The numbers in them — run-rate growth, compute spend, customer concentration — will tell you more about Claude's future than any product announcement. The first one arrives in a few months.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Sources: Anthropic, PBC, Form S-1/A Amendment No. 2 (filed November 10, 2026) and Amendment No. 3 (filed November 17, 2026), SEC EDGAR. Pricing and first-trading dates are reported by CNBC, citing people familiar with the deal, and are not stated in the filing. This article is informational, not investment advice.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>anthropic</category>
      <category>claude</category>
      <category>ai</category>
      <category>ipo</category>
    </item>
    <item>
      <title>Cursor's Security Incident, Explained: A Leaked Vendor Key, a Poisoned Extension, and 5,000 Users at Risk</title>
      <dc:creator>Tiến Thành bình</dc:creator>
      <pubDate>Sat, 14 Nov 2026 02:20:30 +0000</pubDate>
      <link>https://dev.to/tin_thnhbnh_6d6c9102d/cursors-security-incident-explained-a-leaked-vendor-key-a-poisoned-extension-and-5000-users-47h4</link>
      <guid>https://dev.to/tin_thnhbnh_6d6c9102d/cursors-security-incident-explained-a-leaked-vendor-key-a-poisoned-extension-and-5000-users-47h4</guid>
      <description>&lt;p&gt;On November 13, Cursor published the full investigation into a security incident it first disclosed on October 29: a malicious, unauthorized copy of Anthropic's Claude Code extension had appeared in the Cursor VS Code extension marketplace. The post is worth reading in full, because the attack chain is a case study in how a single leaked vendor key can cascade through a supply chain. Here is what actually happened, what was and wasn't affected, and what to do if you were in the exposure window.&lt;/p&gt;

&lt;h2&gt;
  
  
  The attack chain, step by step
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;1. It started at Fireworks.&lt;/strong&gt; An attacker obtained an access key that Cursor had issued to Fireworks, an inference vendor. The key was exposed by Fireworks, and the attacker used it to copy data from a read-only storage bucket Cursor used to share data with the vendor.&lt;/p&gt;

&lt;p&gt;That bucket held limited coding agent traces, mostly from 2025, from sessions that were not in Privacy Mode — data Cursor was permitted to use for model training. A small number of those traces contained credentials that had not been redacted. Cursor says less than 0.01% of customers had API credentials exposed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Stolen credentials opened more doors.&lt;/strong&gt; Starting with those credentials, the attacker reached select SaaS tools used by Cursor. From there they accessed archived, non-production Cursor source code repositories (some containing business contact information), limited customer support data and email conversations, product usage metrics and contact details, summaries of some coding agent conversations, and some credentials users had sent to the Cursor support team.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. The attacker poisoned the extension marketplace.&lt;/strong&gt; On October 29, using a credential stolen from the exposed internal data, the attacker placed malicious, unauthorized copies of Anthropic's Claude Code extension directly on the cache used to serve the Cursor VS Code extension marketplace.&lt;/p&gt;

&lt;p&gt;Two points Cursor stresses: Anthropic did not publish this copy, and this was not a vulnerability in Anthropic's product. The marketplace source, Open VSX, never had the malware. The poisoned build lived only in Cursor's serving cache.&lt;/p&gt;

&lt;p&gt;At most 5,000 users who requested specific cached versions of the extension on October 29 potentially received the malicious build. It was available for approximately 11 hours before Cursor removed it. Cursor took the extension marketplace offline and worked with the cloud provider hosting the attacker's server to take that down too.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Notification.&lt;/strong&gt; On October 30, Cursor published a security notice on its blog and on X, with indicators of compromise, remediation steps, and scripts to check your system, and began notifying affected users and team admins through in-app notices and email. Team admins whose credentials were exposed received a list of the specific keys to rotate, and Cursor asked the third-party services that issued those credentials to revoke them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Containment.&lt;/strong&gt; Cursor says the last attacker activity it has seen was on November 1, and it was blocked. The incident is believed contained, with monitoring continuing.&lt;/p&gt;

&lt;h2&gt;
  
  
  What was not affected
&lt;/h2&gt;

&lt;p&gt;The list of what stayed safe matters as much as the breach itself:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Your prompts and code, if you use Privacy Mode. For users not in Privacy Mode, coding agent traces were exposed for fewer than 0.05% of users — each has been notified directly.&lt;/li&gt;
&lt;li&gt;Any data or live traffic you are sending to Cursor right now. The exposed traces were limited to storage used for historical offline analysis.&lt;/li&gt;
&lt;li&gt;The production systems and infrastructure that run Cursor.&lt;/li&gt;
&lt;li&gt;The Cursor app and its official updates, and the code shipped to clients and cloud services. Cursor ships from a separate code hosting system the attacker never accessed, and no release signing keys were exposed.&lt;/li&gt;
&lt;li&gt;Other products, including Cloud Agents, Grok Bot, and all other SpaceX SI products.&lt;/li&gt;
&lt;li&gt;Any other extension in the marketplace — Cursor validated that all others were authentic, and reviewed and rebuilt the marketplace store.&lt;/li&gt;
&lt;li&gt;Active inference requests, which were not exposed. As a precaution, Cursor moved all inference away from the third-party training and inference vendor.&lt;/li&gt;
&lt;li&gt;Any plugins downloaded from the Cursor Marketplace, and customer code repositories.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What Cursor changed
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Moved all inference traffic off Fireworks, deactivated the key issued to it, and cut off its access to all shared data sets.&lt;/li&gt;
&lt;li&gt;Rotated every exposed Cursor credential found, highest risk first, and as a precaution rotated all credentials in use across its internal systems, even where there was no evidence of exposure.&lt;/li&gt;
&lt;li&gt;Validated all other extensions in the marketplace and rebuilt the store.&lt;/li&gt;
&lt;li&gt;Expanded network access restrictions across service providers, so compromised credentials can't be used outside their intended location.&lt;/li&gt;
&lt;li&gt;Added secret scanning and protections across all code repositories, and expanded scanning to other internal data sources.&lt;/li&gt;
&lt;li&gt;Redeployed all production infrastructure and services as a precaution, reviewing every internal service before restoring it.&lt;/li&gt;
&lt;li&gt;Engaged external forensics experts, and contacted everyone affected.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  If you use Cursor, check this
&lt;/h2&gt;

&lt;p&gt;If you installed or updated to the tampered versions — &lt;strong&gt;Anthropic.claude-code 2.1.312 or 2.1.313&lt;/strong&gt; — from the Cursor VS Code extension marketplace during the exposure window, follow the steps in Cursor's security notice. Uninstalling the extension alone may not be enough; the notice includes indicators of compromise and scripts to check your system.&lt;/p&gt;

&lt;p&gt;If you received no notification from Cursor, your data was not in the exposed traces.&lt;/p&gt;

&lt;h2&gt;
  
  
  The bigger lesson for developers
&lt;/h2&gt;

&lt;p&gt;Three things stand out from this incident:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Vendor keys are blast-radius decisions.&lt;/strong&gt; The breach did not start with Cursor's own infrastructure. It started with a key issued to a third party, for a read-only bucket, that turned out to contain unredacted credentials in training traces. Every API key you hand to a vendor is a trust boundary — scope it, monitor it, and assume the data behind it can leak.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The cache is part of the supply chain.&lt;/strong&gt; The malicious extension never touched Open VSX, the marketplace source. It was injected into the cache that serves the marketplace. If you run any distribution layer — a CDN, a mirror, a cache — that layer is as much a target as the origin, and it needs the same integrity checks.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Privacy Mode is not just a preference.&lt;/strong&gt; The single clearest dividing line in this incident was whether sessions were in Privacy Mode. Traces from Privacy Mode sessions were not in the exposed data at all. For tools that train on your usage by default, the privacy toggle is a security control, not a setting to leave for later.&lt;/p&gt;

&lt;p&gt;Cursor apologized to those impacted and said it will update the post if anything important changes. For a tool that sits between your code and the models that read it, that follow-through is the part worth watching.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Source: Cursor, "An update on our recent security incident," November 13, 2026 — &lt;a href="https://cursor.com/blog/security-incident-update" rel="noopener noreferrer"&gt;https://cursor.com/blog/security-incident-update&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>security</category>
      <category>ai</category>
      <category>claudecode</category>
    </item>
    <item>
      <title>Meta's Muse Ember: Haiku-Level Intelligence at a Third of the Cost — What It Means for Agent Builders</title>
      <dc:creator>Tiến Thành bình</dc:creator>
      <pubDate>Tue, 10 Nov 2026 02:19:56 +0000</pubDate>
      <link>https://dev.to/tin_thnhbnh_6d6c9102d/metas-muse-ember-haiku-level-intelligence-at-a-third-of-the-cost-what-it-means-for-agent-3he9</link>
      <guid>https://dev.to/tin_thnhbnh_6d6c9102d/metas-muse-ember-haiku-level-intelligence-at-a-third-of-the-cost-what-it-means-for-agent-3he9</guid>
      <description>&lt;p&gt;On November 9, 2026, Meta introduced Muse Ember, the smallest and most efficient model in the Muse family. It is not pitched as a frontier model. It is pitched as the model that does the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.&lt;/p&gt;

&lt;p&gt;That framing matters. Most of an agent's token spend is not one brilliant reasoning leap. It is hundreds of small, structured steps. Ember is built for exactly that layer — and Meta's headline claim is that it delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting.&lt;/p&gt;

&lt;h2&gt;
  
  
  The three-tier lineup
&lt;/h2&gt;

&lt;p&gt;Ember completes a division of labor Meta has been building toward:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Muse Wonder&lt;/strong&gt; — for the hardest problems.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Muse Spark&lt;/strong&gt; — for coding and agentic workflows.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Muse Ember&lt;/strong&gt; — for repeatable coding and agentic tasks at scale.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The line Meta itself uses: "Muse Wonder plans, Muse Ember handles the volume." Ember is also the first model trained using Muse Wonder Full, Meta's largest model — the teacher-student pipeline is now part of the product story, not just a training detail.&lt;/p&gt;

&lt;h2&gt;
  
  
  Price and specs
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Muse Ember&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Input&lt;/td&gt;
&lt;td&gt;$0.10 per million tokens&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cached input&lt;/td&gt;
&lt;td&gt;$0.004 per million tokens&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Output&lt;/td&gt;
&lt;td&gt;$0.50 per million tokens&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Context window&lt;/td&gt;
&lt;td&gt;1M tokens&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Max output&lt;/td&gt;
&lt;td&gt;512K tokens&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Model ID&lt;/td&gt;
&lt;td&gt;&lt;code&gt;muse-ember-1.0&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;API&lt;/td&gt;
&lt;td&gt;Responses API, Meta Model API, Muse Code&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;For comparison, Meta lists Muse Spark 1.3 at $1.25 per million input tokens and $4.25 per million output tokens. On Artificial Analysis's cost-per-task measure, Ember comes in at about $0.14 per task versus $2.39 for Spark 1.3 — roughly a 17x gap for a one-point difference on the Intelligence Index (36 vs 37).&lt;/p&gt;

&lt;p&gt;Ember is natively multimodal: text, images, audio, and video in, so agents can read screenshots, charts, and scanned documents, or work from voice memos and meeting recordings without a separate model.&lt;/p&gt;

&lt;h2&gt;
  
  
  The benchmarks: where it wins
&lt;/h2&gt;

&lt;p&gt;All figures below are at maximum reasoning effort, from Meta's official announcement, using Artificial Analysis's November 9, 2026 leaderboard and Meta's own internal runs:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Benchmark&lt;/th&gt;
&lt;th&gt;Muse Ember&lt;/th&gt;
&lt;th&gt;Claude Haiku 5.5&lt;/th&gt;
&lt;th&gt;GPT-6 Luna&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Artificial Analysis Intelligence Index&lt;/td&gt;
&lt;td&gt;36&lt;/td&gt;
&lt;td&gt;36&lt;/td&gt;
&lt;td&gt;28&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AutomationBench-AA (business workflows)&lt;/td&gt;
&lt;td&gt;37.8&lt;/td&gt;
&lt;td&gt;13.9&lt;/td&gt;
&lt;td&gt;20.2&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Agents' Last Exam (end-to-end agent tasks)&lt;/td&gt;
&lt;td&gt;61.6&lt;/td&gt;
&lt;td&gt;57.0&lt;/td&gt;
&lt;td&gt;57.1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Terminal-Bench 4.0 (agentic terminal coding)&lt;/td&gt;
&lt;td&gt;40.9&lt;/td&gt;
&lt;td&gt;39.2&lt;/td&gt;
&lt;td&gt;16.4&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Coding Agent Index-AA (real agent harnesses)&lt;/td&gt;
&lt;td&gt;46&lt;/td&gt;
&lt;td&gt;37&lt;/td&gt;
&lt;td&gt;31&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;OSWorld 2.1 (computer use, partial credit)&lt;/td&gt;
&lt;td&gt;69.5&lt;/td&gt;
&lt;td&gt;68.9&lt;/td&gt;
&lt;td&gt;50.5&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PerceptionBench with tools (visual perception)&lt;/td&gt;
&lt;td&gt;80.6&lt;/td&gt;
&lt;td&gt;78.6&lt;/td&gt;
&lt;td&gt;62.8&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GDP.pdf-AA (professional long documents)&lt;/td&gt;
&lt;td&gt;22.8&lt;/td&gt;
&lt;td&gt;20.8&lt;/td&gt;
&lt;td&gt;22.8&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The headline is the Intelligence Index tie: Ember matches Haiku 5.5's best score at about a third of the cost per task, and beats every Luna setting. Meta also claims Ember sits on the efficiency frontier of the full Artificial Analysis leaderboard — no model scores higher for less per task.&lt;/p&gt;

&lt;p&gt;The agent numbers are the more interesting part. AutomationBench-AA, which runs end-to-end business workflows, is not close: 37.8 against 13.9 for Haiku. That is the benchmark that most resembles what Ember is actually sold for — pulling records, calling tools in sequence, updating systems across apps.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it trails
&lt;/h2&gt;

&lt;p&gt;The same table shows the limits, and Meta published them rather than hiding them. On FrontierCode 2.0-AA — real engineering tasks judged on mergeability — Ember scores 27.2, behind Haiku 5.5's 34.9. On long-document work (GDP.pdf-AA) it ties Luna at 22.8 and only narrowly beats Haiku.&lt;/p&gt;

&lt;p&gt;The pattern is consistent: Ember is strong at executing structured, repeatable agent steps and weaker at the hardest open-ended engineering. That is not a flaw in the pitch; it is the pitch. You do not route your hardest merge to the cheapest model. You route the hundred routine steps around it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Safety and availability
&lt;/h2&gt;

&lt;p&gt;Before release, Meta evaluated Ember under its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity, and loss-of-control risks. Across the categories assessed, the model remained below the framework's high-risk thresholds, and it received the same core safety training and testing as the larger models. Meta also tested how it behaves when acting on a user's behalf — whether it respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages, documents, or tool outputs, and seeks human input before high-impact actions.&lt;/p&gt;

&lt;p&gt;Ember is available today on Muse Code and the Meta Model API as &lt;code&gt;muse-ember-1.0&lt;/code&gt;, and on Google Cloud and Oracle Cloud. Meta says more efficient models are already in training.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it means for agent builders
&lt;/h2&gt;

&lt;p&gt;Three practical takeaways:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Route by step type, not by model loyalty.&lt;/strong&gt; The economical architecture is now explicit in Meta's own lineup: Wonder plans, Spark handles the hard coding workflows, Ember runs the volume. If your agent spends most of its tokens on file reads, test runs, targeted fixes, and extraction, those steps belong on the cheapest model that clears the quality bar — and Ember's benchmark profile says it clears it for exactly those steps.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. The cost math changes what you can afford to retry.&lt;/strong&gt; At $0.14 per Artificial Analysis task versus $2.39 for Spark, verification loops, self-correction passes, and multi-sample checks stop being luxuries. A small model that is cheap enough to run three times can beat a big model run once — provided the task is the kind Ember is good at.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Check the harness before you trust the number.&lt;/strong&gt; Meta is transparent that Terminal-Bench 4.0 ran each model in its best published harness (mini-swe-agent for Ember, Claude Code for Haiku, Codex for Luna), and that some results are Meta's own internal evaluations. The direction of the results is consistent across independent and internal benchmarks, but the exact margins depend on setup. Pilot it on your own workload before migrating volume.&lt;/p&gt;

&lt;p&gt;The bigger signal is structural: the frontier labs are now competing on the cost of the &lt;em&gt;repeated&lt;/em&gt; step, not just the hardest one. Ember is Meta's bet that the agent economy will be won at the bottom of the stack — and at $0.10 per million input tokens, it is a bet other builders can now test for the price of a coffee.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Source: Meta AI Research, "Introducing Our Smallest, Most Efficient Model: Muse Ember," November 9, 2026 — &lt;a href="https://research.meta.ai/blog/muse-ember" rel="noopener noreferrer"&gt;https://research.meta.ai/blog/muse-ember&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>meta</category>
      <category>llm</category>
      <category>agents</category>
    </item>
    <item>
      <title>Cursor Now Ships Your App, Not Just Your Code: What 'Build Live Apps' Actually Changes</title>
      <dc:creator>Tiến Thành bình</dc:creator>
      <pubDate>Fri, 06 Nov 2026 02:18:38 +0000</pubDate>
      <link>https://dev.to/tin_thnhbnh_6d6c9102d/cursor-now-ships-your-app-not-just-your-code-what-build-live-apps-actually-changes-57ie</link>
      <guid>https://dev.to/tin_thnhbnh_6d6c9102d/cursor-now-ships-your-app-not-just-your-code-what-build-live-apps-actually-changes-57ie</guid>
      <description>&lt;p&gt;Agents made writing code cheap. Shipping the result never got cheap with it. That is the gap Cursor is now trying to close: on November 5, 2026, it released "Build Live Apps in Cursor," a workflow that takes a described app from prompt to a hosted, shareable URL — database, auth, and hosting included.&lt;/p&gt;

&lt;p&gt;The interesting question is not whether the demo works. It is what changes when the last mile of deployment becomes something an agent does for you, and where the new failure modes live.&lt;/p&gt;

&lt;h2&gt;
  
  
  What actually happens
&lt;/h2&gt;

&lt;p&gt;The flow, per Cursor's changelog, is deliberately boring:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Kick off a cloud agent and say you want to build an app.&lt;/strong&gt; You have to state it explicitly — this is not something that happens by accident while you are editing code.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The agent scaffolds the web app.&lt;/strong&gt; It writes the application code, spins up a Postgres database if the app needs one, sets up authentication behind Cursor SSO, and provisions hosting. If the app needs to reach your own systems, like a data warehouse, the agent asks for the secrets it needs instead of guessing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;You pick a unique domain.&lt;/strong&gt; Cursor hands back a live link.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;You publish.&lt;/strong&gt; Only after the security checks pass — and only when you explicitly publish — does anything go live.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Apps can be deployed from new or existing repos on Cursor Origin and GitHub. GitLab support is coming later. The capability is available on all paid plans with cloud agents enabled.&lt;/p&gt;

&lt;h2&gt;
  
  
  The publish gate is the real feature
&lt;/h2&gt;

&lt;p&gt;The scaffolding is the part that demos well. The publish gate is the part that decides whether teams can actually use this.&lt;/p&gt;

&lt;p&gt;When you publish, Cursor runs deterministic and agentic security checks designed to catch common issues — secrets written into app files, exposed credentials, that class of mistake. If a check fails, the agent tells you what it needs to fix. Nothing goes live until the checks pass and you explicitly publish.&lt;/p&gt;

&lt;p&gt;Team admins get a second layer of control: they can see every app and secret on the team, add their own rules to the security review, approve publishes the review blocked, and unpublish any app.&lt;/p&gt;

&lt;p&gt;That matters because the failure mode of "anyone can ship an app" is not bad code — it is a well-meaning colleague publishing an internal tool with a hardcoded credential and a public URL. The gate does not eliminate that risk, but it moves the check from "remember to review before deploying" to "deployment cannot happen until the check passes."&lt;/p&gt;

&lt;p&gt;Sharing follows the same logic. Apps can be shared with specific people, groups, or the whole team, all behind Cursor sign-in. On individual plans, apps can also be shared publicly. You can provision a Cursor API key if a script or a bot needs to use your app.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who this is actually for
&lt;/h2&gt;

&lt;p&gt;Cursor's own examples are telling. Inside the company, people across sales, operations, legal, finance, and engineering are building internal apps: event registration and desk seating tools, multimedia presentations for customers, training and onboarding experiences, live sales dashboards, and portals for managing engineering services.&lt;/p&gt;

&lt;p&gt;That list is the honest target market. This is not a replacement for your production stack. It is a replacement for the internal tool that currently lives as a half-finished repo on someone's laptop, or as a spreadsheet with a Slack bot duct-taped to it.&lt;/p&gt;

&lt;p&gt;The pattern to watch: an internal tool used to need a developer, a repo, a deployment pipeline, and someone willing to own it in production. "Build Live Apps" compresses the middle three into a prompt. The remaining question is who owns the app after it ships — and the admin controls suggest Cursor knows that question is coming.&lt;/p&gt;

&lt;h2&gt;
  
  
  The limits, stated plainly
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Paid plans only, with cloud agents enabled.&lt;/strong&gt; If you are on a free plan, this does not change your workflow.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Postgres, SSO, and managed hosting.&lt;/strong&gt; The deployment path fits web apps that work inside that model. Anything needing custom networking, a different identity system, or specialized infrastructure still needs a conventional pipeline.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GitHub and Cursor Origin only, for now.&lt;/strong&gt; GitLab support is announced but not shipped.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;You still have to ask for it.&lt;/strong&gt; The agent builds an app only when you explicitly tell it to. There is no ambient "deploy everything I write" mode, which is a good thing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The security checks are a gate, not a guarantee.&lt;/strong&gt; Deterministic checks catch known-bad patterns; agentic review adds judgment. Neither is a substitute for deciding, as a team, what is allowed to be public.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What to try first
&lt;/h2&gt;

&lt;p&gt;If you have a paid plan with cloud agents, the lowest-risk first app is the one your team already wants but nobody will own: a small internal dashboard, a signup form for an event, a status page for a service. Something with real users inside your org and no secrets worth stealing.&lt;/p&gt;

&lt;p&gt;Watch three things on that first attempt: whether the agent asks for the right secrets instead of inventing them, whether the publish check catches a deliberately planted mistake, and whether an admin can actually see and unpublish the app afterward. Those three behaviors tell you more about whether this fits your team than any demo.&lt;/p&gt;

&lt;p&gt;The localhost graveyard was never a coding problem. It was a deployment-ownership problem. Cursor just automated the deployment half. The ownership half is now the interesting part — and it is still yours.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Source: Cursor changelog, "Build Live Apps in Cursor" (November 5, 2026) — &lt;a href="https://cursor.com/changelog" rel="noopener noreferrer"&gt;https://cursor.com/changelog&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>webdev</category>
      <category>ai</category>
      <category>devops</category>
    </item>
    <item>
      <title>Binance's $6.3B hot wallet breach: what happened, and what users should actually do</title>
      <dc:creator>Tiến Thành bình</dc:creator>
      <pubDate>Sat, 31 Oct 2026 08:46:09 +0000</pubDate>
      <link>https://dev.to/tin_thnhbnh_6d6c9102d/binances-63b-hot-wallet-breach-what-happened-and-what-users-should-actually-do-22m7</link>
      <guid>https://dev.to/tin_thnhbnh_6d6c9102d/binances-63b-hot-wallet-breach-what-happened-and-what-users-should-actually-do-22m7</guid>
      <description>&lt;p&gt;On October 30, 2026, at 04:05 CST, Binance's hot wallets were hit by unauthorized withdrawals totaling about $6.3 billion — the largest crypto exchange theft on record. When a number is that big, the worst response is panic. Facts first, then actions.&lt;/p&gt;

&lt;p&gt;What happened, per Binance's official announcement: the attack involved unauthorized withdrawal transactions from hot wallets. The preliminary investigation points to a DevOps engineer's laptop being compromised, which gave the attacker access to a service container used to submit those transactions. Binance says it has found no evidence so far of wallet private keys being leaked. A full technical postmortem is still pending, and anything beyond the official account — especially claims about who did it — is speculation or, at most, the assessment of chain-analysis firms, not Binance's conclusion.&lt;/p&gt;

&lt;p&gt;The response has three parts, and they are worth keeping separate because they answer different questions.&lt;/p&gt;

&lt;p&gt;First, freeze. Binance says it worked with blockchain partners to freeze about $3 billion of affected assets on-chain, with recovery efforts ongoing. Frozen is not the same as returned, but it means that portion stopped moving.&lt;/p&gt;

&lt;p&gt;Second, cover. Binance says it reserved over $6.34 billion in company funds for this incident, with more than $5 billion worth of corresponding tokens already in place — USDC, USDT, BTC, ETH, BNB, among others — roughly 80% of the stolen amount. The rest is being acquired through market and over-the-counter channels, which takes longer for less liquid tokens.&lt;/p&gt;

&lt;p&gt;Third, restore. Binance says it will fully cover the losses through its Secure Asset Fund for Users (SAFU) and company funds, resume withdrawals in phases within about 34 hours, take a first proof-of-reserves snapshot within 12 hours, and publish the results by November 2. That proof of reserves is the public yardstick: it is how anyone outside Binance can check whether compensation actually landed, rather than taking a press statement on faith.&lt;/p&gt;

&lt;p&gt;Why "a big exchange got hacked" does not automatically mean "the money is gone": the breach itself is severe — an attack path that bypassed internal controls well enough to drain hot wallets at this scale is not something to downplay. But what users ultimately lose depends on whether the freeze, cover, and restore steps actually execute, and on the proof of reserves that follows. Watch those, not just the day-one headline.&lt;/p&gt;

&lt;p&gt;What ordinary users should do now:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;Do not panic-move. While withdrawals are paused, any link or "support agent" promising fast-track or insider withdrawals is almost certainly phishing. Resumption happens through official announcements only, never through private messages or small community accounts.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Verify your sources. Read Binance's official announcements and official accounts. Third-party retellings of amounts, timelines, and causes should be discounted until the official postmortem is out, and attacker attribution in particular should be treated as unverified.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Diversify custody. Do not keep everything on one platform. This is the one habit worth making permanent after an event like this: separate funds you trade with from funds you hold long term, and keep a buffer across hot wallets, cold wallets, and different platforms — not by moving frantically, but as a standing arrangement.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Check your own security settings now, while nothing is on fire. Is two-factor authentication on? If you changed phones, did your authenticator migrate properly? Are your recovery email and phone number ones you still control? Are there login devices you do not recognize? These checks take minutes when done calmly, and they are miserable to do after an incident.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The news cycle on this will pass. What does not expire is the boring part: diversified custody, working security settings, and discipline about sources. Platforms can fail; your response should not have only one layer.&lt;/p&gt;

</description>
      <category>security</category>
      <category>crypto</category>
      <category>news</category>
    </item>
  </channel>
</rss>
