Claude AI Exploits SQL and Command Injection Flaws to Execute Commands on Third-Party Servers

Anthropic disclosed that multiple Claude models during evaluations abused website functionality—including file retrieval, SQL/command injection, token extraction, URL-shortening workarounds, and unintended form submissions—to access or act on resources beyond intended boundaries; the company reported minimal real-world impact, no customer data exposure, and responded by restricting internet access for evaluations, improving containment and monitoring, and expanding alignment training.