OpenCode Breach Exposes AI API Keys and Triggers Charges Above $1,000
OpenCode says attackers exposed Console API keys, triggering charges above $1,000 and exposing gaps in AI models used for cyber defense for users worldwide.
Summary
OpenCode said attackers breached an older system overnight Saturday, exposing API keys stored in OpenCode Console workspaces for 1.4% of platform users. Developers reported unauthorized OpenAI, Anthropic and Google AI spending, often hundreds of dollars and sometimes above $1,000. Anthropic and Google blocked suspected keys; OpenAI helped restore access and refund consumed credits; OpenCode notified every affected user, began refunds and ordered provider key replacement. Co-founder Dax Raad said the open-source app and command-line tool, which never receive those keys, were unaffected.
Attackers kept targeting OpenCode for 48 hours, forcing round-the-clock AI defense. Models initially helped, then refused malicious-code analysis when safety filters activated; access to OpenAI’s Daybreak Blue program, which offers vetted users models less likely to reject defensive work, improved the investigation. OpenCode called the breach a large-scale distillation attack. Leaked keys could generate the high-volume outputs needed for distillation where American models are restricted, but skeptics say that use is unproven. OpenCode has not explained entry, key storage or total misuse; the incident remains ongoing.
The breach coincided with OpenCode making GLM-5.3-Flash free; researchers say the GLM-5.3 family’s cyber capability approached Anthropic’s restricted Claude Mythos Preview. In July, Hugging Face faced similar refusals while analyzing an autonomous-agent intrusion that OpenAI later identified as its own lowered-safeguard test. David Sacks, co-chair of the President’s Council of Advisors on Science and Technology, said guardrails backfired. Earlier this month, CrowdStrike linked attacks on South Korean banks to a likely China-based hacker using an AI penetration-testing agent and Claude; last week, India’s cybersecurity agency warned that AI agents entered Indian organizations without authorization.
Positives
- Anthropic and Google blocked API keys believed to have been exposed through OpenCode Console.
- OpenAI restored customer access and is refunding API credits consumed during the attack.
- OpenCode notified all affected users, began refunds and instructed them to revoke and replace provider keys.
- OpenAI’s Daybreak Blue access improved OpenCode’s ability to investigate malicious code without defensive requests being rejected.
- OpenCode’s open-source app and command-line tool were unaffected because their API keys never reach the company.
Risks & concerns
- 1.4% of OpenCode Console users had stored API keys exposed through an older system.
- Individual users reported unauthorized AI charges in the hundreds of dollars, with some totals exceeding $1,000.
- Attackers targeted OpenCode continuously for 48 hours while its defensive AI models began refusing malicious-code analysis.
- OpenCode has not disclosed the intrusion method, key storage design or full extent of misuse, and the incident remains ongoing.
- CrowdStrike and India’s cybersecurity agency recently flagged separate AI-enabled intrusions affecting South Korean banks and Indian organizations.