Cursor has explained how attackers got a malicious Claude Code extension into its marketplace: through agent logs stored with its vendor Fireworks that contained credentials in plain text. For any team giving agents access, that's the real lesson.
Key facts
- According to Cursor's November 13, 2026 report, an access key exposed by Fireworks AI was the entry point; attackers copied agent traces from non-Privacy-Mode sessions, some containing unredacted credentials.
- On October 29, malicious versions 2.1.312 and 2.1.313 of the Claude Code extension sat in the marketplace cache for about 11 hours; up to 5,000 users may have received them.
- Cursor says fewer than 0.01% of customers had API credentials exposed; it has moved all inference traffic off Fireworks.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
You know the move. The agent asks for an API key, you paste it into the chat, the task runs, done. The key doesn't go away. It sits in the log. And sometimes that log lives with a vendor you've never heard of. At Cursor, that exact chain turned into a backdoor that may have reached up to 5,000 users.
What happened
On November 13, Cursor, the AI code editor, published its incident report. The entry point was an access key Cursor had issued to Fireworks AI, the vendor it used for model inference and training. According to Cursor, Fireworks exposed that key. Attackers used it to copy data from a storage bucket Cursor shared with Fireworks: coding-agent traces, mostly from 2025, from sessions without Privacy Mode, which Cursor was allowed to use for training.
The sentence that matters: some of those traces contained unredacted credentials. With them, the attackers got into select SaaS tools Cursor uses: archived code repositories, support emails, usage data, and credentials users had sent to support. On October 29 they placed two malicious copies of Anthropic's Claude Code extension (versions 2.1.312 and 2.1.313) in the cache of Cursor's extension marketplace for about eleven hours. BleepingComputer reports the payload tried to install a persistent SSH backdoor. Cursor has since moved all inference traffic off Fireworks.
Fireworks should sound familiar: two days ago I wrote about the Notion leak at the same vendor. The new part isn't the name. It's what the data could do.
A chat log isn't an archive anymore. It's a key ring.
At Notion, conversations leaked. Ugly, but the data stayed data. At Cursor, the data opened doors. That's the difference between a chatbot and an agent: an agent works with real access. Whatever you hand it so it can act ends up in the transcript. And the transcript travels further when it's cleared for training.
To me that's the real wake-up call, and it isn't only for developers. Marketing teams are handing agents access to ad accounts, shops, CRMs and newsletter tools right now. The more an agent can do, the more its log is worth. A key pasted into a chat stops being a convenience and becomes a master key with an unknown retention period.
AI didn't create the problem. It scaled it.
A password in an email to support is an old mistake. What's new is the volume: agents record every step, teams paste keys by the minute, and the logs end up with vendors who have vendors of their own. BleepingComputer says Wiz describes a whole wave of attacks on AI developer platforms in October. AI doesn't fix sloppy credential habits. It makes them scale.
What I'd do now
- Stop pasting keys into chats. Access belongs in the tool's settings or integrations, not in the conversation history.
- Rotate every key that has ever appeared in an AI chat or a support ticket this week. It takes an hour.
- For every agent tool your team uses, find out: do transcripts go into training, and to whom? That setting is a decision, not a default you inherit.
Bottom lineGive an agent a key, and you give it to everyone who reads the log later.
Got AI news? Submit news →
Source
Cursor: Security incident update
More sources
- BleepingComputer – 14.11.2026, SSH-Hintertür, Versionen, Wiz „Black October“
- Cursor – Hinweis vom 29./30.10.2026 mit Prüf- und Bereinigungsschritten




