Update: Cursor's backdoored extension started with a Fireworks AI key and API keys left in agent logs
Cursor's root-cause report: a key exposed in the Fireworks AI breach opened coding-agent traces with unredacted credentials, which led to the tampered Claude Code extension.
This is an update to our November 6 report, Tampered Claude Code extension on Cursor's marketplace planted an SSH backdoor, and connects to our coverage of the Fireworks AI breach.
On November 13, 2026, Cursor published the root cause of its October incident: an access key Cursor had given to its AI vendor Fireworks AI was exposed in Fireworks' breach. That key opened a bucket of old coding-agent session logs, some of which still contained users' credentials. Those credentials led to Cursor's internal tools and, eventually, to the tampered Claude Code extension. Cursor says it has directly notified affected users. If you were notified, rotate the keys it lists. If you installed Claude Code 2.1.312 or 2.1.313 from Cursor's marketplace on October 29, follow the cleanup steps; uninstalling is not enough.
What happened, in plain English
Cursor is an AI code editor. When developers let its AI agent work on their code, the editor records a "trace": a log of what the agent saw and did. That can include terminal output, configuration files and anything the developer pasted in. Unless a user turns on Cursor's Privacy Mode, Cursor is allowed to use these traces to train its models. Cursor shared some of them with Fireworks AI, an outside company it used for AI training and for running models ("inference"), through a read-only cloud storage bucket.
In October, Fireworks AI disclosed a breach. According to Cursor, the attacker found among the stolen material the key Cursor had issued to Fireworks, and used it to copy that bucket. Most of the traces were from 2025. Some of them contained passwords and API keys (the secret tokens programs use to log in to services) that had not been blanked out, or "redacted".
Each step opened the next one. The vendor breach exposed a key, the key exposed the logs, and the logs exposed more credentials. With those, the attacker got into some of the business software Cursor uses. They then used a stolen credential to plant a modified copy of Anthropic's Claude Code extension in the storage that serves Cursor's extension marketplace. That copy was available for about 11 hours on October 29.
Think of it as a spare office key left with a cleaning contractor. The contractor is burgled, the thief finds the key, and in your office they find a notebook with more passwords written in it.
Are you affected? What to do now
Most people have nothing new to do. Cursor says the incident did not affect users with Privacy Mode turned on, customer code repositories, live traffic or active AI requests, its production systems, the Cursor app and its official updates or signing keys, or any other marketplace extension.
You are in scope if any of these apply:
- You got a notification from Cursor saying your credentials or your agent traces were exposed. Cursor puts the exposed API credentials at fewer than 0.01% of customers. It puts exposed traces at fewer than 0.05% of users who were not using Privacy Mode. Action: revoke and rotate every key listed in the notice now. Check the provider's logs for use you don't recognize since October.
- You ever sent a password, token or key to Cursor support. Cursor lists "credentials users had sent to Cursor support" among the data the attacker accessed. Action: rotate those credentials even if you were not notified.
- You installed or auto-updated Anthropic.claude-code 2.1.312 or 2.1.313 from the Cursor marketplace on October 29, 2026. Action: follow Cursor's October 29 security notice. Check
~/.ssh/authorized_keysfor the attacker's key and look for the persistence items it lists. Remove them or rebuild the machine, then rotate any credentials that machine could reach. In Cursor's words, "Uninstalling the extension alone may not be enough." Our November 6 article has the full checklist and the published indicators of compromise.
Cursor also says the attacker saw archived non-production code, limited support emails, product usage metrics, contact details and summaries of some agent conversations. If you have dealt with Cursor support, be wary of follow-up phishing that uses those details.
A wider lesson for every team using AI coding tools. Any agent session can capture a secret: a .env file it read, a token printed in a terminal, a key you pasted. Where your tool offers a no-retention or privacy setting, use it for work involving credentials. Keep secrets out of agent context where you can. Run secret scanning on logs and exports, not only on Git repositories.
The expert view
Root cause. Cursor's report describes a chain of three trust failures, and no single software bug. First, a credential handed to a third-party AI vendor became a single point of failure once that vendor was breached. Second, the data behind that credential was meant for model training, yet it carried live secrets. Cursor says some traces "contained credentials that had not been redacted". It does not say how its redaction works or why it missed them. Third, those secrets gave access to SaaS tools, and from there to the extension cache. The credential that opened the cache is not named.
Why this is an AI-specific problem. Agent traces are richer than ordinary chat logs. An agent reads files and runs commands, so its record of a session holds whatever was in the developer's environment: config files, environment variables, command output. Training pipelines have a reason to keep this data for a long time and share it with partners. Most of these traces were from 2025, which shows how long that exposure can last. This is the same pattern as the Notion AI transcripts exposed through Fireworks. When a company sends AI data to a vendor for training or inference, a breach at that vendor also exposes the company's own users.
Attack chain, conceptually. The steps were vendor compromise, a reused partner credential, bulk data theft, harvested secrets, access to SaaS tools, and finally write access to a distribution cache. The last step turned a data breach into a supply-chain attack on developer workstations. Each step used legitimate credentials. That explains why the attack worked, and also why Cursor's monitoring could spot it as unusual activity rather than as an exploit.
Severity in context. No CVE or CVSS score applies. By Cursor's figures, the share of users whose credentials or traces were exposed is very small. But the credentials were real and the payload was a working SSH backdoor, so for affected users the impact is high. What makes this important is not the number of victims. It is that the incident shows a realistic, repeatable path from an AI vendor's breach into the customers' own machines.
What Cursor changed. Cursor says it moved all inference off Fireworks and cut the vendor's access to its data. It rotated every exposed credential it found, plus all internal credentials as a precaution. It also added secret scanning across its repositories and other internal data, tightened network restrictions for service providers "where possible", rebuilt the marketplace and redeployed production. It reports no attacker activity after November 1.
Still unknown. Cursor does not say how long agent traces are kept or what is redacted from them today. It has not published a timeline for when the key was taken from Fireworks, and it does not name the credential that opened the marketplace cache. Cursor did not attribute the attack to any group, and we have not seen an attribution from another credible source. Fireworks has not, in the material we reviewed, commented on Cursor's account.
Official sources
- Cursor — Security incident update, November 13, 2026
- Cursor — Security notice: malicious Anthropic.claude-code (2.1.312 and 2.1.313), October 29, 2026 (updated October 30)
- Fireworks AI — October 2026 security incident update
- Coverage: The Hacker News — Stolen Fireworks AI key led to Cursor breach (November 15, 2026)
- Coverage: Cloud Security Alliance — Signed by the right publisher: trojanized Claude Code on Cursor's marketplace