OpenCode Console breach exposed stored OpenAI, Anthropic and Google API keys, and OpenCode ties it to a distillation attack

The breach hit an old Console system that OpenCode was already shutting down. That system kept teams' linked provider keys on OpenCode's servers. OpenCode has not said the local app's keys, which sit in a file on the user's own machine, were involved. It also has not said how the attacker got in, how many accounts the 1.4% represents, or whether anyone will be reimbursed. [4] [8] [5]

OpenCode called the incident a distillation attack, meaning the keys were used to pull frontier-model outputs at scale while victims paid. It has published no logs to back that label. A few days earlier, Anthropic said its new defenses cut the reasoning tokens distillers extract by roughly 90–99% and that it often bans such accounts within two hours. Those figures are Anthropic's own estimates from its own platform. One explanation is that real, trusted customer keys become more valuable as fake accounts get caught faster. That explanation is a theory and has not been confirmed. [4] [6] [5]

Anthropic's research suggests why full reasoning text is the prize. In an internal experiment, a model trained on full Claude reasoning transcripts scored 80.4% on SWE-bench Verified. A model trained only on summarized reasoning scored 71.8%. The experiment has not been released for others to reproduce. [6] [7]

Affected users are advised to revoke and reissue keys in each provider's dashboard, update any apps that used them, check usage from October 16 onward, and file a support ticket for fraudulent charges. More broadly: use one key per tool, cap spending on every key, and inventory old or retired tools that may still hold live keys. On Anthropic, custom limits are set per workspace and cannot be applied to the default workspace. [3] [5]