LegalPwn: Tricking LLMs by burying badness in lawyerly fine print

Pangea researchers describe "LegalPwn," a novel prompt-injection technique that embeds adversarial instructions within legal text so that LLMs and agentic tools ingesting those documents follow the hidden instructions. In testing, several models and tools (including gemini-cli and GitHub Copilot) were tricked into misclassifying malicious code as safe and, in live scenarios, recommending or executing reverse shells; mitigations proposed include input validation, contextual sandboxing, adversarial training, human review, and vendor fixes.