LLMs Are a New Type of Insider Adversary

This commentary warns that while LLMs are becoming vital business tools, they introduce serious risks such as jailbreaks, prompt-driven manipulation, and remote code execution—risks amplified when models are integrated with sensitive systems. It cites research (e.g., 31% of targeted codebases exhibiting LLM-induced RCE and a LangChain flaw enabling reverse shells), argues current guardrails and tools like content filters and Llama Guard don’t address root causes, and urges an 'assume breach' posture. Recommended mitigations include enforcing least privilege, not relying on LLMs as a security perimeter, limiting model actions by impersonating users, sanitizing data and outputs, and sandboxing code, with the OWASP Top 10 offering further guidance.