Cursor’s Incident Shows Why Old Logs Need a Credential Review

A read-only archive can still contain a key that opens another service. Cursor’s incident shows why old AI agent records deserve a credential review.

Cursor’s Incident Shows Why Old Logs Need a Credential Review

An old troubleshooting conversation can outlive the problem it solved. If it contains a credential that still works, whoever can read that conversation may also be able to use the account behind it. Cursor’s latest incident report is a useful reason to look beyond the permissions on an archive and ask what access its contents could unlock.

How historical data became a route into other systems

In its November 13 update, Cursor described the chain behind an incident disclosed in October. According to the company, an attacker obtained a key exposed at Fireworks and copied historical coding-agent records, or traces, from a read-only storage bucket. These records were mostly from 2025 sessions outside Privacy Mode. Some contained unredacted credentials, which helped the attacker reach internal services.

A further stolen credential enabled malicious copies of the Claude Code extension to be placed in Cursor’s extension-distribution cache on October 29. Cursor says Anthropic did not publish them, Open VSX did not host the malware, and its own application and official updates were unaffected. The company says it believes the incident is contained.

Fireworks’ November 9 bulletin supplies related context: it reports impacts involving secrets and, for certain customers, datasets and incomplete models, alongside account and billing information. Fireworks also says its investigation remains ongoing and that it has not detected impact to its Inference, Training SDK or Nexus products.

Cursor says it contacted affected customers directly. It says fewer than 0.05% of users had coding-agent traces exposed, and those traces did not include Privacy Mode sessions. That distinction concerns the stored traces, not protection from the separate malicious extension.

If you used the affected extension

Cursor’s remediation notice identifies a specific exposure: Windows or macOS installations or updates of Anthropic.claude-code versions 2.1.312 or 2.1.313 from its VS Code Extension Marketplace between approximately 2:27 a.m. and 1:35 p.m. Pacific on October 29.

If that matches your installation history, follow Cursor’s detection and remediation instructions. Removing the extension alone may leave persistence mechanisms behind. The company advises isolating hosts with evidence of execution or a backdoor, assessing whether they need rebuilding, and rotating accessible credentials and sessions from a trusted system. This article has not tested those remediation scripts.

The question to ask about your own archives

The broader lesson is that storage permissions and credential permissions answer different questions. A person might be unable to edit a saved transcript while still being able to copy a key from it. That key is then evaluated by another service, under that service’s rules. Calling the transcript “read-only” says little about the power of the credential inside.

OWASP’s secrets-management guidance gives teams a practical starting point: restrict each credential to the access it needs, prefer short-lived credentials where supported, and revoke secrets that are no longer needed or may be compromised. It also recommends recording what a secret is used for so it can be found and managed.

For a small project, start with one integration. Can you identify its key’s owner, scope, expiration and replacement process? Could you revoke it without guessing which application would break? If you discover a live key in a shared log, removing that copy does not invalidate copies someone may already have taken. Address the credential at the issuing service, then review how it entered the log.

Keeping useful debugging history has real value. The decision should include how long embedded access remains useful to someone else.

Sources