AI Agent Exploitation, Prompt Injection, and Vulnerability Discovery Acceleration

An unpatched critical unauthenticated LMCache RCE creates urgent risk for exposed or root-running AI infrastructure, while PoeLLM reportedly compromised more than 3,400 exposed servers. Anthropic's reported free scans and broader AI-offense claims indicate accelerating defensive and offensive experimentation, but validation, coverage, exploitability, and disclosure governance remain unresolved; MCP OAuth weaknesses and prompt injection add identity and workflow risk.