Cursor's breach started with API keys left unredacted in coding agent traces it shared with a vendor. Scrub secrets before logs leave
Cursor published its investigation into the October breach, and the entry point was agent trace data that still contained credentials.
What happened
Cursor says an attacker obtained an access key that Cursor had issued to Fireworks, its training and inference vendor. The key was exposed on the Fireworks side, and the attacker used it to copy a read-only storage bucket Cursor used to share data. That bucket held coding agent traces, mostly from 2025, from sessions not in Privacy Mode. A small number of those traces contained credentials that had not been redacted, and Cursor says fewer than 0.01% of customers had API credentials exposed.
From those credentials the attacker reached some of the SaaS tools Cursor uses, then archived non-production repositories, some support data and email, usage metrics, and credentials users had sent to Cursor support. On October 29, using a stolen credential, the attacker placed tampered copies of Anthropic's Claude Code extension, versions 2.1.312 and 2.1.313, on the cache behind Cursor's extension marketplace. Cursor's earlier notice says the build tried to add an attacker SSH key and open a persistent reverse tunnel. At most 5,000 users may have received it during the roughly 11 hours it was live.
Cursor says it moved all inference off Fireworks, rotated every credential across internal systems, restricted where provider credentials can be used from, and added secret scanning to its repositories and other internal data sources. It says the last attacker activity was on November 1 and was blocked.
My take
The part I'd act on is the first link in the chain. Agent traces capture everything the agent saw, including environment variables, config files and the API key someone pasted into a chat. Once those traces are copied to an analytics store or a vendor bucket, the keys travel with them.
In the automations I build, the same risk sits in execution logs. An n8n or Make run log can hold a full HubSpot token or a webhook secret if a node echoes headers. I treat those logs as sensitive data: redact secrets before they are stored, keep retention short, and never ship raw runs to a third party for debugging or training.
Cursor's network restriction step is also worth copying. A key that only works from your own server's IP is far less useful to whoever steals it. If you installed either tampered version, follow Cursor's notice, since it says uninstalling alone may not be enough.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
