OpenCode Says Attackers Used AI Agents To Expose Console API Keys, Impacting 1.4% Of Users
Technology and ScienceUpdated 1 hour ago2 min read
Keys exposed through an old system
OpenCode said attackers broke into an old computer system that was being shut down. This exposed the API keys of some OpenCode Console users. The company linked the incident to a large-scale attack involving a method called distillation, which can use the answers from one AI model to train another. The breach affected 1.4% of OpenCode Console users, and OpenCode said it emailed those affected. Co-founder Dax Raad said attackers had been targeting OpenCode's systems with AI agents for at least 48 hours. He said the situation became stable after OpenCode began using OpenAI's Daybreak AI cybersecurity model to help defend its systems.
OpenCode told affected users that their keys might have been exposed as early as October 17, 2026, at 00:30 UTC. The company said there was no sign that other data had been exposed. BleepingComputer reported that some users saw charges they did not recognize and found that their projects in Google Cloud had been suspended after the breach.
Charges, key replacements, and the size of the breach
OpenCode said the breach involved API keys from other companies that had been saved in affected workspaces. These keys let users connect to outside services. OpenCode told users to cancel the exposed keys and create new ones using each provider's dashboard. The company said that anyone who did not receive an email had not been affected. It also said the incident was still ongoing when it published its statement. Co-founder Dax Raad wrote, "Note this is the OpenCode Console, not OpenCode itself, those keys never come to us," as users reported unexpected charges on their OpenAI, Anthropic, and Google AI accounts.
A developer posted on Reddit and X that alerts about spending on OpenAI, Anthropic, and Google AI had "all fire off before going to bed wondering what was going on." Other users reported unexpected charges worth hundreds of dollars, and a few said their totals were above $1,000. OpenCode said it was working with Anthropic, OpenAI, and Google to fix the problem. It also said refunds for charges users had not approved were being processed.
Limits of the defense and what happens next
OpenCode said the attackers used the stolen keys to make requests that cost money. BleepingComputer described the incident as a distillation attack in which attackers gathered large amounts of information by sending queries without permission. OpenCode said it was working with Anthropic, OpenAI, and Google to fix the problem. It told affected users to cancel and replace their keys right away and carefully check their recent bills for API requests. TechCrunch reported that OpenCode's investigation ran into problems when the AI models it was using began refusing requests because of their built-in safety rules. Dax Raad said, "[T]hat was a bit frustrating," as these safety rules blocked some efforts to investigate and defend against the attack.
David Sacks wrote on X that OpenCode and an earlier incident involving Hugging Face "hit the same wall mid-attack: frontier models refused to analyze the very exploits needed for defense because the payloads tripped safety filters." He said the safety rules had the opposite effect from what was intended. OpenCode's public statement did not name an attacker or provide technical evidence showing how the breach supported the reported distillation activity. It also did not explain how the attackers gained access or how many keys were involved.