OpenCode breach exposes OpenAI API keys for 1.4% of Console users in "distillation attack"

#What happened

OpenCode, the open-source terminal coding agent built by Anomaly Innovations (the company formerly known as SST, and a different product from OpenAI's Codex), disclosed on October 17 that attackers gained unauthorized access to a legacy system it was in the process of decommissioning. The intrusion may have exposed third-party provider API keys, including OpenAI, Anthropic and Google keys, saved in affected users' workspaces. OpenCode described the incident as a "large-scale distillation attack" that impacted roughly 1.4% of OpenCode Console users, and says it emailed every affected user the morning of October 17. If a user did not receive that email, the company says they were not impacted.

#What we know

• The attack targeted a legacy Console system being retired; OpenCode has not said how the attackers got in or how the keys were stored, citing an ongoing incident.
• Affected developers reported unauthorized charges on their OpenAI and Anthropic bills before the notification arrived, and at least one user reported a Google Cloud project suspended for abuse.
• OpenCode co-founder Dax Raad said his personal accounts were also being attacked.
• The company says it is working with Anthropic, OpenAI and Google on remediation, and has asked affected users to revoke and replace keys through each provider's dashboard, update downstream tools, and review recent usage.

#Why it matters

This is a structural warning for the "bring your own key" AI-tool economy: OpenCode is provider-agnostic, so one breach of a single third-party console exposed OpenAI, Anthropic and Google credentials at once. Beyond billing risk, OpenCode's "distillation attack" framing matters: provider keys used this way can turn victims' paid model access into a conduit for harvesting model outputs. OpenCode is unrelated to OpenAI's Codex, but it is where many OpenAI API keys live, which is why OpenAI is named as one of the providers coordinating on the response.

Verified Oct 17, 2026