Two weeks after a malicious copy of Anthropic's Claude Code extension was discovered lurking in its extension marketplace, AI coding company Cursor has published the most detailed account yet of how the attack happened — and the first domino, according to the company, was an access key issued to AI inference vendor Fireworks AI that Fireworks exposed.

In a blog post dated November 13 titled "An update on our recent security incident," Cursor laid out an investigation that connects a leaked vendor credential, unredacted training data and a poisoned marketplace cache into a single attack chain. The company says the incident is now contained, but the disclosure is a case study in how third-party AI vendor relationships can become security liabilities — an issue that touches every team building with AI tools today.

The First Domino: An Exposed Vendor Key

Cursor's investigation began, in the company's telling, at Fireworks AI — the inference provider it used for model serving and training data sharing. An attacker obtained an access key that Cursor had issued to Fireworks, and that key "was exposed by Fireworks," according to Cursor's account.

With the key, the attacker copied data from a read-only storage bucket that Cursor used to share data with the vendor. That bucket contained limited coding agent traces — mostly from 2025, drawn from sessions not conducted in Privacy Mode, which Cursor was permitted to use for model training. The critical failure was inside those traces: a small number contained credentials that had not been redacted. Cursor says fewer than 0.01% of its customers had API credentials exposed as a result.

From those stolen credentials, the attacker pivoted further inward, reaching select SaaS tools used by Cursor and, from there, accessing select archived, non-production source code repositories. Cursor emphasizes that its production systems and the infrastructure that runs its product were not accessed.

From Stolen Credentials to a Poisoned Marketplace

The attack's most alarming phase came on October 29. Using a stolen credential surfaced from the exposed data, the attacker placed malicious, unauthorized copies of Anthropic's Claude Code extension — versions 2.1.312 and 2.1.313 — directly onto the cache used to serve Cursor's VS Code extension marketplace.

Cursor is explicit about where responsibility does not lie: Anthropic did not publish the copies, and this was not a vulnerability in Anthropic's product. The upstream marketplace source, Open VSX, never carried the malware — the tampering happened in the cache layer that Cursor operates.

The exposure window was short but real. At most 5,000 users who requested specific cached versions of the extension on October 29 potentially received the malicious build, and it was available for approximately 11 hours before Cursor removed it. The company then took the entire extension marketplace offline while it reviewed the damage.

Detection, Containment and Cutoff

According to the post, Cursor's monitoring systems identified the unauthorized activity during the attack, prompting a rolling response: the vendor's key was deactivated, an incident response workstream opened, and credentials began being revoked and rotated while the team hunted for further access. As new access was identified, it was cut off — and, out of caution, Cursor temporarily disabled internal and product services well beyond what the evidence strictly required, which disrupted some features before they were restored.

The company says the last attacker activity it observed was on November 1, and it was blocked. Cursor believes the incident is contained, though it continues to monitor for anomalous activity. External forensics experts have been engaged to support the investigation.

What Was Exposed — and What Wasn't

The blast radius, per Cursor's own accounting, is narrow but not negligible:

  • Coding agent traces were exposed for fewer than 0.05% of users who did not use Privacy Mode. Cursor says it has identified these users and notified each of them directly.
  • API credentials held by fewer than 0.01% of customers appeared in unredacted traces. Team admins received emails listing the specific keys to rotate, and Cursor asked the third-party services that issued those credentials to revoke them.
  • Prompts and code in Privacy Mode were not exposed. Nor were live traffic, active inference requests, the production systems that run Cursor, the official app and its update pipeline — which ship from a separate code hosting system the attacker never touched — or any release signing keys.
  • All other extensions in the marketplace were validated as authentic, and the marketplace store itself was reviewed and rebuilt.

Cursor Cuts Fireworks Loose

The remediation list makes clear that the vendor relationship did not survive the incident. Cursor has moved all inference traffic off Fireworks, deactivated the key it had issued to the company, and cut Fireworks' access to every dataset Cursor shared with it — an effective termination of the training-data arrangement, which other reporting on the incident characterized as Cursor dropping the vendor outright.

Beyond the cutoff, Cursor says it has rotated every exposed credential it found (and, as a precaution, all credentials in use across its internal systems), expanded network access restrictions so that compromised credentials cannot be used outside their intended locations, added secret scanning across its code repositories and other internal data sources, and redeployed all production infrastructure as a precaution.

The Supply-Chain Lesson

The post-mortem is likely to be read closely across the AI tools industry, because it documents a chain of failures that any company sharing data with AI vendors could replicate: credentials embedded in training traces that escaped redaction, a vendor-side key exposure, and a marketplace cache trusted as if it were upstream source.

The practical takeaways are already familiar to security teams but bear repeating in an AI context. Keys issued to third parties should be scoped, short-lived and monitored. Data shared for training should be treated as likely to leak, and redaction should be verified rather than assumed — especially for conversational traces, where developers habitually paste secrets. And distribution infrastructure — caches, mirrors, marketplaces — needs the same integrity monitoring as source code, because it is exactly where a small, time-boxed tamper can reach thousands of developers silently.

For the affected users, Cursor's October 30 security notice remains the operative guidance: check whether the tampered versions of the Claude Code extension were installed during the exposure window, run the provided detection scripts, and rotate any credentials the company has flagged. The company has apologized to those impacted and says it will update its post if anything material changes.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →