OpenCode Console breached, stored API keys exposed

A bad weekend. OpenCode, the open-source AI coding agent, announced a security breach from its X account on Saturday, 17 October. The target was OpenCode Console. Console is the hosted service where teams store their OpenAI, Anthropic or Google key and authenticate their model requests with it. According to the company, the attackers got into an old system that was being retired. The earliest moment the keys may have been exposed is 17 October, 00:30 UTC.

Those affected make up 1.4 percent of Console users. Only users who had connected an outside provider are in that group, and the company told them by e-mail. According to co-founder Dax Raad, the open-source app and the CLI (the command-line tool) were not affected, because keys kept locally never reach the company in the first place. OpenCode calls the incident a "distillation" attack (using hijacked access to harvest model output and train other models on it). That claim has not been independently confirmed.

Users reported bills of several hundred dollars run up overnight. In a few cases the amount went past $1,000, and in one report it came close to $1,700. Anthropic and Google blocked the keys they believe were exposed. OpenAI, for its part, said it would refund the credits spent in the attack.

My advice is simple: deleting the key inside OpenCode isn't enough, because that does nothing to a copy someone has already taken. Revoke and replace the key with each provider separately, and go through your bills going back to 16 October.