OpenCode cloud platform breached: third-party API keys suspected to have been leaked, users report being charged $400

Beating AI News: AI coding tool OpenCode has disclosed a security incident affecting 1.4% of its Console users. All affected users had linked third-party model services via the cloud. Public notifications from users show an old system slated for deactivation was accessed without authorization, potentially exposing OpenAI, Anthropic, and other API keys stored in workspaces. The OpenCode team has sent emails to affected users. OpenCode offers both local coding tools and an optional cloud-based Console service. Users can entrust their API keys to the Console for it to call models on their behalf, allowing team members to share these models without accessing the original keys directly. Once attackers obtain these keys, they can bypass user login to directly call paid models, with charges billed to the original account holders. A Reddit user reported $400 in unauthorized calls to their Anthropic account. OpenCode linked the incident to "large-scale model distillation attacks" — a practice of collecting responses from one model in batches to train another model with that data. If attackers steal a large number of API keys, they may use others' accounts to collect training data, shifting high model call costs onto the original users. OpenCode is collaborating with Anthropic, OpenAI, and Google to resolve the incident. Anthropic's Lydia Hallie stated the company has revoked potentially affected Claude API keys. OpenCode is urging affected users to visit the respective model service providers' backends to revoke and replace their keys, and check for recent abnormal calls.