Live keys in a system already being retired

OpenCode is a coding agent with about 215,000 GitHub stars, and the company says 16 million developers use it every month. On Saturday it said a breach had hit a legacy system it was in the process of decommissioning. Keys saved in user workspaces may have been exposed. The affected share of Console users, the vendor said, was 1.4%. It gave no absolute count.

The Console is not the coding CLI. By the vendor's account, only keys linked to the Console were exposed. Users reported charges. One reported $281 on OpenAI and $168 on Anthropic, starting at 2 a.m. ET. Another put his losses at $1,700. A third saw more than $700 of OpenAI usage before he pulled the plug. The Latent could not independently verify the figures.

In a Saturday post, the vendor said the breach was related to a large scale distillation attack. Distillation trains one model on another's outputs. The term describes how stolen keys might be used, not how they were obtained. The Latent reported that OpenCode did not explain how it linked the leaked keys to the attack.

Nathan Lambert asked whether the vendor meant 'distillation' as used to train downstream models, or 'some other weird shady pattern to use illegally obtained tokens.' Another user was blunter: 'This is not a distillation attack.' The cause is contested.

The breach is real, and the user reports say some of the keys worked. What caused it is not settled. The vendor's own account put the breach in a system it was already retiring, which is the last place credentials should be.

Log take: A breach disclosure that names an attack technique before it shows the evidence invites the suspicion that it is managing the story. (Interpretation; contested framing, rests on the vendor's account.)