After OpenCode and Railway, Command Code also attacked: hackers target AI model keys

Beating AI News Flash, October 29 - AI programming tool Command Code announced it has prevented a security incident related to a large-scale distillation attack targeting Anthropic and OpenAI models. The company proactively revoked all of its own API keys, and users need to log in again to generate new keys.

Over the past two weeks, OpenCode and Railway suffered similar attacks in succession. On October 17, OpenCode's cloud Console was breached, affecting approximately 1.4% of Console users, with some third-party model keys potentially leaked, and users reported unauthorized charges of hundreds or even thousands of dollars.

On October 21, Railway was also compromised, with up to approximately 100,000 workspaces at risk of environment variable leakage. Investigators found from the attackers' code logs that the hackers were specifically searching for AI service API keys. Railway determined that the attackers most likely intended to use these keys to make large-scale model calls and collect outputs for distillation.

Even earlier, on October 14, Zeabur was also breached. The attackers focused on searching for AI API keys and even attempted to use stolen AWS credentials to call Amazon Bedrock, but were unsuccessful.