The motive wasn't the money — it was a 'distillation attack' to clone the models

The motive in the OpenCode breach was not the money, the company says, but a distillation attack: intruders who broke into a legacy Console system used the stored API keys to fire a huge volume of queries at OpenAI, Anthropic and Google models, capturing the responses to train a cheaper model that mimics the originals.

The legacy system was already being decommissioned, but it still held live credentials belonging to 1.4% of Console users. Reported losses vary by account: one user saw $281 in unexpected OpenAI charges and $168 on Anthropic, another more than $700 overnight, a third above $1,700. All three providers have revoked the compromised keys, and refunds are underway.

The pattern is the point: tools built to hold spend-capable credentials make a single leak expensive. Per-key spend limits, and a separate scoped key for each tool, would have capped the damage — a project-level ceiling set at a fraction of $1,700 would have ended that user's night in minutes.