AI news · OpenCode · security
OpenCode tells users to replace API keys after a breach
The company behind the coding agent says intruders reached an old system where some users’ OpenAI and Anthropic keys were stored.
OpenCode, a coding agent that works inside a developer’s terminal, told some of its users on Saturday 17 October 2026 to revoke their AI provider keys. An API key works like a password that lets a program run, and bill, your OpenAI or Anthropic account. Anomaly Innovations, the company behind OpenCode, says intruders got into a legacy system, one that is being shut down, Startup Fortune reports.
About 1.4 per cent of OpenCode Console users were hit, the company says. Only workspaces with an outside provider’s key saved in them were exposed. Anomaly ties the breach to a large-scale distillation attack. In distillation, a lab copies a model’s skills by feeding a second model huge numbers of its replies, and every reply is a paid query. Anomaly has not said how the stolen keys and the distillation connect, and some researchers, Nathan Lambert among them, have questioned the label.
The notice asks users to swap their keys and look for usage they do not recognise, Runtime Wire reports. Anthropic, OpenAI and Google are helping with the response, Anomaly says. Some users report charges for queries they never made, Startup Fortune adds, and the company says its inquiry has found no other data taken so far.
If your team keeps model keys inside a third-party tool, a monthly cap on every key puts a ceiling on what a leaked one can spend.
Sources
- Startup Fortune: OpenCode breach exposes stored OpenAI and Anthropic keys, prompts urgent key rotation. Checked
- Runtime Wire: OpenCode warns users to rotate provider keys after legacy-system intrusion. Checked
More AI news from
- Meta drops a terms clause critics read as a licence to scrape
- DeepSeek updates its Flash model and opens its search engine
- ChatGPT ads can target users by email and phone, 404 Media finds
- Disney sets $180 million for 66 Character.AI hires
- Big websites shut the door on AI agents, The Verge finds
- PostgreSQL 19 drops features after AI checks find defects
- Replit makes OpenAI its default model over Anthropic
- Waymo plans robotaxi rides in Toronto from 2028
- Google now pays over 400 publications for AI news features