The Risks of Code Assistant LLMs: Harmful Content, Misuse and Deception

This Palo Alto Networks report examines security risks in LLM-based coding assistants, focusing on indirect prompt injection via attached external context, misuse of auto-completion to generate harmful content, and direct model invocation or stolen-session misuse. The authors demonstrate simulated scenarios where contaminated context leads the assistant to insert an obfuscated backdoor that fetches and executes commands from an attacker-controlled C2, discuss auto-complete jailbreaks and LLMJacking risks, and provide mitigations such as manual review, execution control, and tighter context handling.