Microsoft finds Claude Code bug that could leak secrets

Microsoft researchers found a security hole in Anthropic's Claude Code GitHub automation. Attackers could use prompt injection to steal secret keys from CI/CD workflows.

Prompt injection is an AI vulnerability where attackers hide malicious instructions in content to trick the model. Microsoft gave an example where instructions were hidden in HTML comments, invisible on GitHub but readable by AI models. Attackers don't need project edit rights, just a GitHub ticket to make the AI bot do bad things.

Microsoft reported the bug to Anthropic on April 29. Anthropic fixed it on May 5 with Claude Code version 2.1.128, blocking access to sensitive files in the /proc/ directory.