A Contractor's Key Let a Fake Claude Code into Cursor's Marketplace: Up to 5,000 People in 11 Hours
Cursor says an attacker planted a fake Claude Code from Anthropic in its extension marketplace. It started with a storage key that Cursor gave to its contractor Fireworks. The key leaked, and old agent logs contained unredacted credentials. The fake stayed up for about 11 hours, and up to 5,000 people could have downloaded it. Strip tokens from logs before handing them to contractors.
AI-processed from Cursor Blog; edited by Hamidun News
Imagine you gave a contractor the key to a data warehouse, they accidentally exposed it, and the data inside included old AI agent session records with other people's passwords in them. That is how the incident Cursor disclosed on October 29 began, according to the company's own account. A malicious copy of Anthropic's Claude Code appeared in the extension store for its editor.
The fake stayed up for about 11 hours, and at most 5,000 people could have downloaded it. Fireworks is an external vendor that runs models (inference) and takes part in training. Privacy Mode is a Cursor setting under which a user's code and prompts are not stored for training.
If you build agents or share logs with vendors, this story is about you.
How the chain started with a contractor
According to Cursor, it all began at Fireworks. Cursor had issued it an access key to a storage bucket where the company published data for joint work. Fireworks exposed that key, and the attacker copied the contents of the bucket, which was open for reading only.
The bucket held limited traces of coding agent sessions, mostly from 2025. A trace is a record of a session: what the user asked, what the agent read, and what it replied. It included sessions from people who had not enabled Privacy Mode, which meant they had allowed their data to be used for training. A small share of the traces still contained unredacted credentials: API keys that people pasted into the chat or that the agent read from a project. Fewer than 0.01% of customers were affected.
With those keys, the attacker gained access to several SaaS services Cursor uses. From there, they reached archived repositories (not production code, partly containing business contacts), limited support data and correspondence, usage metrics, and contacts. They also obtained short summaries of some agent conversations and keys that users had sent to support. Cursor's monitoring systems flagged the unusual activity. The company disabled the vendor's key, began rotating credentials, and searched for other traces of access.
A fake extension in the marketplace cache
On October 29, the attacker used a key stolen from an internal system to place malicious copies of the Claude Code extension directly into the cache from which Cursor serves extensions. Anthropic had not published this version, and its product had no vulnerability. The malicious code was also not present in Open VSX, the source the marketplace draws extensions from.
Only those who requested the specific cached versions that day could have been infected, at most 5,000 people. Anthropic.claude-code 2.1.312 and 2.1.313 are named as the dangerous versions. Cursor removed the fake, took the marketplace offline, and, together with its cloud provider, got the attacker's server shut down. On October 30 the company published a bulletin with indicators of compromise and verification scripts, then notified affected users in the app and by email. Team admins received lists of keys that need to be replaced.
If you installed or updated these versions, removing the extension may not be enough. Follow the steps in the security bulletin.
What is unaffected and what is changing
Cursor says that prompts and code from users with Privacy Mode enabled were not affected. Among users who did not use the mode, agent traces reached the attacker for fewer than 0.05% of users, and the company notified each of them directly. Production infrastructure, the application itself and its updates (which are built in a separate system where release signing keys were not exposed), live traffic, customer repositories, Cloud Agents, and other extensions were not affected. Cursor re-verified the other extensions and rebuilt the marketplace storage.
Among the protective measures, the company listed the following:
- All inference has been moved off Fireworks, and the key issued to it has been deactivated;
- All leaked keys found were replaced, and then, as a precaution, all internal keys;
- Network restrictions were tightened so a stolen key cannot be used outside its intended location;
- Secret scanning is enabled across all repositories and other internal sources;
- All production infrastructure was redeployed from scratch, and external experts were brought in for the investigation.
The attacker's last activity was noticed on November 1 and was blocked. Cursor believes the incident has been contained.
What to take away
First, agent logs and traces have turned into a store of secrets: people and the agents themselves routinely read and display .env files and tokens. If such data goes to a contractor or into a training set, it has to be scrubbed before handoff, not after.
Second, vendor keys should be narrow: read-only, scoped to the specific bucket, bound to addresses, and short-lived. Third, extension distribution caches need the same scrutiny as the original catalog: the substitution happened there, not at the source.
There is a caveat: this is the company's self-report. The text contains no independent verification, and the number of affected people is given as an upper bound. The details are specific enough to check your own setup: look at which Claude Code versions you have, and rotate any keys you ever sent to support or pasted into an agent outside Privacy Mode.
- Nov 21, 2026Anthropic Raises $81B in IPO at $2 Trillion Valuation as Protesters Show Up Outside Its Offices
- Nov 21, 2026Banks Stand to Earn $550 Million on Anthropic's IPO, More Than on SpaceX
- Nov 21, 2026Anthropic's IPO Order Book Is Several Times Larger Than the Shares on Offer; Founders Set to Receive $1.75 Billion
- Nov 21, 2026Anthropic Researcher Claims He Proved the Riemann Hypothesis with Claude in Six Weeks
Key Questions Answered
How the chain started with a contractor?
According to Cursor, it all began at Fireworks. Cursor had issued it an access key to a storage bucket where the company published data for joint work. Fireworks exposed that key, and the attacker copied the contents of the bucket, which was open for reading only. The bucket held limited traces of coding agent sessions, mostly from 2025. A trace is a record of a session: what the user asked, what the agent read, and what it replied.
What are the key details on A fake extension in the marketplace cache?
On October 29, the attacker used a key stolen from an internal system to place malicious copies of the Claude Code extension directly into the cache from which Cursor serves extensions. Anthropic had not published this version, and its product had no vulnerability. The malicious code was also not present in Open VSX, the source the marketplace draws extensions from. Only those who requested the specific cached versions that day could have been infected, at most 5,000 people.
What is unaffected and what is changing?
Cursor says that prompts and code from users with Privacy Mode enabled were not affected. Among users who did not use the mode, agent traces reached the attacker for fewer than 0.05% of users, and the company notified each of them directly. Production infrastructure, the application itself and its updates (which are built in a separate system where release signing keys were not exposed), live traffic, customer repositories, Cloud Agents, and other extensions were not affected.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.