OpenCode breach exposed provider API keys — rotate yours now

A retired Console system leaked keys for 1.4% of users; stolen ones rang up four-figure bills and froze Google Cloud projects. What to do, and how it happened.

Nowline OCT 20 6:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • What leaked — and whose keys

    OpenCode — the open-source terminal coding agent — says its cloud Console, not the CLI, exposed linked provider API keys for 1.4% of Console users. OpenAI, Anthropic, and Google keys are in scope; earliest exposure was Oct 17, 00:30 UTC. Co-founder Dax Raad says the keys never touched OpenCode itself.

  • The bill came due

    Attackers used the stolen keys to run billable requests. Affected users report unauthorized charges from the hundreds to over $1,000 across OpenAI, Anthropic, and Google accounts, with some Google Cloud projects suspended and spend alerts firing overnight. OpenCode says refunds are underway.

  • Rotate from the provider, not OpenCode

    If you ever linked a provider to OpenCode Console, revoke and reissue those keys in each provider's own dashboard — not via OpenCode or any link in a breach email — then audit billing from Oct 16 on and set hard spend caps. No email from OpenCode means you weren't affected.

  • 48 hours of agents on a dead system

    The way in was a legacy Console system OpenCode was decommissioning. Raad says attackers hit it for at least 48 hours using AI agents in what OpenCode calls a "large-scale distillation attack" — stolen keys driving unauthorized queries. No attacker has been named, and the incident was still live at disclosure.

  • The twist: safety filters slowed the defense

    Reportedly, OpenCode's own AI models refused to help analyze the attack, citing safety safeguards, which stalled the response; things reportedly stabilized once the team switched to OpenAI's Daybreak security model. A live reminder to vet which models you reach for during an incident.

  • Also today: Claude Code 2.1.304

    Anthropic shipped Claude Code 2.1.304 with code-review levels (Quick, Standard, Deep) and an option to review every repo in a GitHub org; /auto-mode-setup now runs on Bedrock, Vertex, and Foundry. Plus 43 bug fixes.