OpenCode Cloud Breached: Third-Party API Keys Suspected Leaked, Users Report $400 Deducted

Beating AI News Flash: AI coding tool OpenCode discloses a security incident affecting 1.4% of Console users. All affected individuals had previously linked third-party model services in the cloud. A public user notification shows that a legacy system being decommissioned suffered unauthorized access, and API keys stored in the workspace, including those for OpenAI and Anthropic, may have been leaked. The team has sent emails to affected users.

OpenCode not only offers a local coding tool but also an optional cloud Console service. Users can entrust their API keys to the Console for hosting, which then calls models on their behalf. Team members can also share these models without directly accessing the original keys.

Once attackers obtain these keys, they can bypass user login and directly call paid models, with the costs still borne by the original users. A Reddit user reported $400 in unauthorized usage charges on their Anthropic account.

OpenCode links the incident to "large-scale model distillation attacks." Distillation involves collecting a model's responses in bulk and using that data to train another model. If attackers steal a large number of API keys, they could use others' accounts to collect training data, shifting the high costs of model calls onto users.

OpenCode is handling the incident with Anthropic, OpenAI, and Google. Anthropic's Lydia Hallie stated that the company has revoked Claude API keys deemed potentially affected. OpenCode has asked affected users to revoke and replace their keys in the respective model service providers' backends and to check for any recent anomalous calls.