← CYBERSPACE / Pwning

How a Stolen AI Vendor Key Put a Backdoor in 5,000 Developers' Editors

hacking

On 29 October 2026, tampered builds of Anthropic's Claude Code extension sat in Cursor's marketplace for about eleven hours. The trail led back through AI infrastructure vendor Fireworks AI to a single leaked access key — and an SSH backdoor on up to 5,000 developer machines. Cursor's full incident report, published 13 November, lays out a textbook software supply-chain attack.

The short version

  • On 29 October 2026, a tampered build of Anthropic's Claude Code extension was served from Cursor's VS Code extension marketplace under the legitimate extension ID anthropic.claude-code — versions 2.1.312 and 2.1.313.
  • It stayed live for roughly eleven hours (about 09:27–20:35 UTC) and reached at most 5,000 users.
  • The payload was a 66-line JavaScript loader that ran when the extension started, planted an attacker-controlled SSH key labelled svc-telemetry, opened a reverse SSH tunnel to 134.209.82[.]225, installed persistence, and sent host details home over plain HTTP.
  • Cursor's two-week investigation, published on 13 November, traced the entire chain back to one access key that leaked in a breach at AI infrastructure vendor Fireworks AI.

Stage 1 — a leaked key at Fireworks AI

The starting point was not Cursor at all. In early October, Fireworks AI — a provider of infrastructure for model training and inference — disclosed a security incident. Its update of 9 November widened the scope: a financially motivated criminal group had reached Fireworks development machines through a third-party cloud provider, then accessed resource metadata, billing information and — for a limited set of accounts — secrets, datasets and incomplete model weights. Fireworks names the motive as "LLMJacking": reselling unauthorised access to large language models to illicit routing gateways.

Cursor had issued Fireworks an access key. That key was caught up in the Fireworks breach — and the attackers used it to copy data out of a read-only storage bucket through which Cursor shared material with the vendor. The bucket held historical coding-agent traces, mostly from 2025, from sessions that were not in Privacy Mode and whose data Cursor was permitted to use for model training.

Stage 2 — from traces to Cursor's internal tools

A small number of those traces contained credentials that had not been redacted. Cursor puts the exposure at fewer than 0.05% of users for traces and fewer than 0.01% of customers for API credentials. Small percentages — but enough. Using those credentials, the attackers reached selected SaaS tools used by Cursor and pulled:

  • 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;
  • credentials that users had previously sent to Cursor's support team.

Cursor's monitoring caught the unauthorised activity; the company revoked the Fireworks key and began rotating credentials. By then, however, the attackers already held one more credential — this time one that could publish extensions.

Stage 3 — poisoning the marketplace

On 29 October, the attackers used that credential to place malicious copies of Anthropic's Claude Code extension directly on the cache that serves Cursor's extension marketplace. This is the detail that makes the incident unusual: the compromise sat in the distribution layer, not in Anthropic's product. The package identified itself correctly as anthropic.claude-code, carried a plausible version number, and came from the editor's own built-in marketplace. Anthropic's systems were not compromised, and the malicious build never appeared on Open VSX, Microsoft's Visual Studio Marketplace, or npm.

The tampered extension was available for about eleven hours, from roughly 02:27 to 13:35 Pacific time. Cursor took the whole marketplace offline, removed the builds and rebuilt the service; the attackers' command-and-control server was taken offline on 30 October with the hosting provider's help.

What the backdoor did

Microsoft's Threat Intelligence team analysed the injected code. The loader was prepended to the extension's main JavaScript bundle and ran with no user interaction required. It:

  1. collected host reconnaissance — hostname, username, OS, Git email, public IP, and the names of up to 40 environment variables matching patterns such as KEY, TOKEN, SECRET, PASS, AWS and GCP;
  2. appended an attacker SSH public key to ~/.ssh/authorized_keys, labelled svc-telemetry to look harmless;
  3. tried to enable the local SSH service and open a reverse tunnel back to the attacker, retrying roughly every 60 seconds;
  4. installed persistence per platform — a macOS LaunchAgent (com.apple.sysmond), a Windows scheduled task (SysMon) plus a Startup entry, and a Linux @reboot cron entry;
  5. sent the host details to 134.209.82[.]225:8099/beacon over unencrypted HTTP.

Two caveats matter for defenders. Enabling local SSH requires administrator rights, which the extension did not have on most Windows and macOS hosts — so the tunnel step often failed. But uninstalling or updating the extension does not remove what it left behind, because the artifacts were written outside the extension directory.

Who is affected — and how to check

The affected window is specific: Cursor users on macOS or Windows who installed or updated anthropic.claude-code from Cursor's marketplace between roughly 09:27 and 20:35 UTC on 29 October 2026. A version number alone proves nothing, because a genuine 2.1.312 also exists on Open VSX.

The practical checks, as published by Cursor, Microsoft and StepSecurity:

  • search extension directories such as anthropic.claude-code-2.1.312-* and anthropic.claude-code-2.1.313-* for extension.js files containing svc-telemetry or the address 134.209.82.225;
  • look for the svc-telemetry key in authorized_keys, for the listed persistence entries, and for a running reverse-tunnel process;
  • treat any host with those artifacts as compromised, isolate it, and rotate every credential reachable from it from a known-clean machine;
  • hunt network logs for traffic to the command-and-control IP.

Why this one matters

Attacks on the supply chain of developer tooling are not new, but this one compresses several modern risks into a single chain. The first loss was not a password — it was a long-lived access key held by a third party — and the pivot ran through data shared for model training: unredacted secrets sitting in coding-agent traces that nobody thought of as a credential store. The final step then exploited the one signal every developer is trained to trust — a verified publisher and a legitimate extension ID in a built-in marketplace.

AI coding assistants raise the stakes, because they run with the developer's own privileges and sit in one of the most credential-dense environments in any company: source trees, cloud keys, CI tokens, signing material. Cursor's own framing is the right lesson — publisher-identity controls, verification badges and account security do not protect the distribution layer that sits between a real publisher and a user. The recommendations that follow are the ones package registries already learned: install cooldowns, extension allowlists and provenance checks for the IDE extension tier.

Sources

View source ↗