Cursor has traced the poisoned Claude Code extension that appeared in its marketplace on 29 October to a single leaked vendor key, closing a two-week investigation into one of the year’s most troubling supply-chain attacks on developers, and publishing a full account of what the attacker reached.
How one key opened the door
The attack began at Fireworks AI, the inference provider Cursor used for model training. An access key Cursor had issued to the vendor was exposed on Fireworks’ side, and the attacker used it to copy data from a read-only storage bucket the two companies shared.
That bucket held limited coding-agent traces, mostly from 2025, drawn from sessions where users had not enabled Privacy Mode and whose data Cursor was allowed to use for training. Some of those traces contained credentials that had never been redacted, and fewer than 0.05% of users had traces exposed.
Starting with those credentials, the attacker reached several of Cursor’s software tools, then took archived non-production source code, limited support data and email conversations, usage metrics and contact details, plus summaries of some coding-agent chats and credentials users had sent to support.
The poisoned extension
On 29 October the attacker used a stolen credential to place malicious copies of Anthropic’s Claude Code extension, versions 2.1.312 and 2.1.313, directly on the cache serving Cursor’s extension marketplace. The builds stayed live for about eleven hours, and up to 5,000 users may have received one.
Cursor stresses that Anthropic did not publish the copies and that no flaw in Anthropic’s product was involved. The marketplace source, Open VSX, never carried the malware. The tampered code collected host details, added an attacker SSH key, and tried to open a reverse tunnel with persistence.
Researchers found coding errors that likely stopped the backdoor working on many Windows and macOS machines where SSH was not already enabled. Even so, Cursor warns that removing the extension alone may not be enough, and affected users should follow its remediation scripts and rotate exposed credentials.
What Cursor changed
Cursor says the incident is contained: the last attacker activity it saw was on 1 November, and it was blocked. The company has moved all inference traffic off Fireworks, deactivated the vendor’s key, rotated exposed credentials and then all internal ones, and rebuilt its extension marketplace from scratch.
Our Read
The striking part is how ordinary the failure was. No zero-day opened Cursor’s doors; a shared bucket, a vendor key and unredacted credentials in training traces did the job. The lesson is about data hygiene as much as access control: secrets hiding in logs are live credentials, not dead history.
The marketplace cache is the other warning. Attackers did not need to compromise Open VSX or Anthropic; poisoning the delivery cache was enough to reach thousands of developers. Extension stores that serve cached builds need integrity checks as the registries they mirror, or the cache becomes the weakest link.
Why It Matters
Coding agents now hold the keys to real systems: API tokens, cloud credentials and repository access. When those agents run on shared infrastructure, one vendor’s breach can hand an attacker the means to poison tools developers trust. Agent traces should be treated as secrets, not telemetry.
Background
We first reported the poisoned extension on 31 October, when Cursor disclosed the malicious builds. The wider Fireworks breach behind it deepened on 9 November, when the vendor said customer datasets and model weights were also exposed, and Notion then confirmed some users’ AI chat transcripts had leaked.
Source: Cursor’s incident update and BleepingComputer’s report.


0 Comments