← CYBERSPACE / Pwning

Anthropic's Week: AI Agents Inside the Building, and a Backdoor in the Developer Tooling

hacking

Two attacks in one week, neither one using a flaw in Anthropic's own code. Cathedral's agents walked in through third-party sales and HR software in under 12 hours, reaching Slack messages, transcripts and salary data. Four days later the official Claude Code extension auto-updated a remote access trojan onto ~5,000 developer machines under Anthropic's own verified account - deleted quietly, and still installed on the machines that ran it.

Two attacks in one week, both aimed at the same company, and neither one needed a zero-day in Anthropic's own software. One came through a vendor's admin panel. The other came through Anthropic's own developer extension, which auto-updated a remote access trojan onto roughly 5,000 developer machines under its verified publisher account.

Part 1: the AI agents that walked in through the side door

Cathedral is a cybersecurity startup founded this year by four former staffers of DOGE, the US government efficiency agency — Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein. It sells AI-powered security tooling, mostly for military and government customers, is valued at 1.4 billion dollars, and joined Anthropic's bug bounty program on 26 October.

What it did next is the part worth reading twice. According to a draft report reviewed by The Washington Post, Cathedral's AI agents went looking for flaws not in Anthropic's own systems, but in the third-party business software Anthropic's sales team ran on. Within about twelve hours the agents had paths to internal employee conversations — including messages about the company's advanced models — plus meeting transcripts, salary data, open job listings and customer contact records.

Anthropic's account is narrower on every axis. It confirms access to a "limited set" of sales, HR and IT systems: customer and partner business contacts, basic user account records. No Claude conversations, no source code, no model weights. It says the attempts stretched over several days, not twelve hours, and that its monitoring flagged some of Cathedral's activity before Cathedral reported it, that it shut off the access the same day, and that it disabled the vulnerable third-party software. It also says Cathedral went "substantially beyond" the bounty's scope, and it has not said whether the 10,000-dollar reward will be paid.

One quote from the reporting has been travelling since, from an unnamed senior Trump administration official: "If a bunch of kids who have just started a company can do this, then potentially the Chinese aren't far behind."

Note what the agents never attacked: the model, the weights, the chat logs. They attacked the boring layer — the SaaS admin panels, the sales tools, the HR stack. That is not a failure of AI security. It is a failure of ordinary vendor management, found by a machine moving at machine speed.

Part 2: the backdoor that shipped inside the official extension

Four days later, the tooling around Anthropic got hit from the other side — and this one was worse.

Version 2.1.313 of the official anthropic.claude-code extension for VS Code-compatible editors was published carrying an SSH reverse backdoor: a remote access trojan, per the security researcher Kevin Beaumont, who documented the incident and published network indicators for it. It shipped under the extension's verified publisher account — no lookalike name, no typosquat — and it arrived on developer machines through the auto-update, not because anyone chose to install it. Cursor dates its window in the Cursor marketplace to roughly 2:27 a.m. to 1:35 p.m. PT on 29 October and the number of affected systems to around 5,000. Phorion, a macOS security firm, was first to spot the malicious activity on endpoints.

What the payload did, per Cursor's advisory, Phorion's analysis and Beaumont's writeup:

  • A malicious IIFE was prepended to the minified extension.js, executing as soon as Node loads the file — before the extension's normal activation. No user interaction required.
  • It collected host and user information: hostname, username, OS, architecture, home directory, environment variable names, working directories, configured Git email, public IP.
  • It appended an attacker-controlled SSH key to the user's authorized_keys, tried to enable the local SSH service and hold open a reverse SSH tunnel to a fixed address — then made that tunnel persistent with a LaunchAgent on macOS, a cron entry on Linux and a scheduled task on Windows with a Startup-folder fallback, retrying roughly every 60 seconds.
  • It beacons host details and backdoor status to the operator's server.

Cursor says at most 5,000 users were affected, that it pulled the version, shut the whole marketplace, audited other extensions, rotated keys, rebuilt it and brought it back, and that it worked with the cloud provider to take the attacker's server offline. It also notes, fairly, that the privileged steps would fail on most macOS and Windows hosts unless the editor was started with administrator rights. The catch is the same as before: the SSH key was still written, the persistence still installed, the machine details still exfiltrated.

Three details from Beaumont's writeup weigh more than the advisory does.

The fix is not a fix. Version 2.1.314 appeared roughly eight hours after 2.1.313 and appears to remove the backdoor from the extension — but it does not remove the SSH access from machines that already ran the malicious build, and it does not undo the reconnaissance. Whoever auto-updated into 2.1.313 is still carrying what it installed. 2.1.313 itself is now absent from Open VSX, the Visual Studio Marketplace and Cursor.

Nobody was told. Beaumont's charge is that Anthropic did not notify customers and that the malicious version was quietly deleted. Cursor's advisory opens with "This malicious version of the extension was not developed by Anthropic" — true in the narrow sense that Anthropic did not author the backdoor, and misleading in the sense that matters: the package was uploaded through Anthropic's own verified publishing account. The supply chain was attacked at the publisher, not in a marketplace side alley.

The blast radius is not five thousand hobbyists. The extension is used by tens of millions of developers, and the organisations running Claude Code and Cursor on laptops include Binance, among thousands of others. Developer laptops hold production credentials, source access and SSH keys — the payload's own shopping list is a good summary of what is worth stealing from one.

How to hunt it

Concrete artifacts, from Beaumont's writeup and Phorion's analysis:

  • Reverse SSH tunnel to 134.209.82.225 port 22; C2 beacon as an HTTP POST to http://134.209.82.225:8099/beacon; plus a loopback attempt to 127.0.0.1:8099/beacon.
  • VSIX SHA-256 (darwin-arm64): 432ccdee3801f99e52b8fe31f69ea9870118e96ea7ef9ecabed1479621b5ab62
  • Modified VSIX SHA-256 (win32-x64): 3de1722901165769893bef57c72594e9a54459faa23f3091b136b79966edd081
  • Public IP lookups to ifconfig.me and api.ipify.org — legitimate services, so the lookup alone is not an indicator.
  • Beaumont published KQL queries for Microsoft Defender for Endpoint to hunt the version across a fleet.

On the endpoint, look for what the malware had to leave behind: new entries in authorized_keys, a LaunchAgent nobody wrote, a cron entry that fires at reboot, a scheduled task named like a system component, and an outbound connection that re-establishes every minute.

What this week says about security

Your perimeter is your vendors, whether you like it or not. Cathedral did not beat Anthropic. It beat whatever sales and HR tooling Anthropic had granted access to. Almost every company in 2026 is one OAuth grant away from the same story — and a bug bounty scope written for your own domains does not cover a vendor's admin console.

Agents turn opportunistic testing into a steady state. A human pentester works office hours. A swarm of agents probes in parallel around the clock at negligible cost, and pivots when it hits a wall — which is exactly what happened here: days against the hardened target, then a turn toward the softer third-party surface. The infrastructure for that was a weekend project in 2024 and is a product today.

The developer workstation is production. The poisoned extension is the second half of the lesson. It reached 5,000 developer machines with a persistent tunnel, through a channel — extension marketplaces — where trust is granted by publisher identity and refreshed by an update. If your build pipeline runs on those machines, a backdoor in the editor is a backdoor in the pipeline.

Detection beats prevention, because the prevention already failed twice. In both incidents the interesting signal was behavioural: an SSH key appearing in authorized_keys, a LaunchAgent or cron entry nobody wrote, an outbound tunnel that reconnects every minute. That is monitorable. The vendor's admin login at 3 a.m. is monitorable. Granting is cheap; watching is what costs.

Scoping machine testers is unsolved. Anthropic and Cathedral now disagree about duration, penetration depth and whether the rules were broken — with the reward money as leverage. When the tester is an agent that does not read the scope document, "in scope" stops being a concept and becomes a legal argument.

There is an irony worth stating plainly. Anthropic is one of the labs that has warned hardest and loudest about where this technology is going. This week it got probed on both flanks by exactly the thing it warns about — once by agents someone pointed at it, and once by a supply chain that ships to everyone. Neither attack touched a single model weight. Both were ordinary security failures, at extraordinary speed — and the second one put a remote access trojan on developer laptops that their owners learned about from a researcher's blog post, not from their vendor.

Sources: https://www.washingtonpost.com/technology/2026/10/26/researchers-cybersecurity-start-up-cathedral-got-inside-anthropic/ https://the-decoder.com/ai-agents-hacked-anthropic-maker-of-arguably-the-most-capable-ai-models-for-cyberdefense/ https://www.independent.co.uk/tech/security/ai-hack-anthropic-cathedral-breach-b3073909.html https://gizmodo.com/former-doge-staffers-new-startup-claims-it-hacked-anthropic-2000830560 https://cursor.com/blog/claude-code-extension-security-notice https://www.helpnetsecurity.com/2026/10/30/claude-code-extension-ssh-backdoor https://thehackernews.com/2026/10/official-claude-code-vs-code-extension.html https://www.darxxun.de/post/the-dam-broke-php-fell-europe-went-dark-and-the-blackwall-arrived-on-schedule https://doublepulsar.com/anthropic-hacked-to-deliver-remote-access-trojan-to-claude-developer-laptops-worldwide-1148a92baae2 (read as archived on 30 Oct 2026)

Keep your authorized_keys short.

View source ↗