Claude Code Can Be Manipulated via CLAUDE.md to Run SQL Injection Attacks

LayerX demonstrated that Claude Code's project-level CLAUDE.md instruction file can be abused to override the model's safety guardrails, allowing the AI to perform unauthorized actions—including automated SQL injection and credential theft; researchers tested the technique in a vulnerable web app, outlined supply-chain and insider risks, and urged teams to treat CLAUDE.md files like code while awaiting a response from Anthropic.