Hackers hit AI coding tool OpenCode, running up charges on users’ AI accounts
The part that got me is the AI refusing to look. Here's a coding agent built to lean on language models for everything, and when the actual attack code lands in front of those same models, they balk. That's a practical limit on how much AI can help with security work, not a theoretical one. Models are trained to avoid analyzing potentially harmful code, and that guardrail kicks in exactly when a developer needs a second pair of eyes on something suspicious.
The API key theft itself is more familiar. These agents run with broad permissions by design, because they need to reach out to services and bill the work. That's the same attack surface that has been hitting other AI coding tools. The real question is how OpenCode handled key storage and rotation, and whether any rate limiting caught the runaway charges before users noticed.
I'd want to know how many keys were exposed and how quickly the team revoked them. Open source tools rely on community trust, and a breach like this is a reminder that convenience and security pull in opposite directions for AI agents.