Friendly Fire — AI Agent Hijacking

Friendly Fire is a proof-of-concept attack published by the Ai Now Institute demonstrating that autonomous AI coding agents (Claude Code and Openai Codex) can be hijacked to execute attacker code during vulnerability scanning of untrusted third-party code.

The Attack

1. An AI coding agent is asked to scan or review a repository for vulnerabilities
2. The repository contains maliciously crafted code that exploits the agent's autonomy
3. The agent executes the attacker's code instead of safely analysing it
4. The attacker gains code execution on the agent's host

Implications

• AI coding agents' autonomy becomes an attack surface
• Reviewing untrusted code with autonomous agents creates a new risk
• Distinction between "reviewing" and "executing" code is blurred for AI agents (raw/digests/Cyber-Digest-2026-07-10)

Related Pages

• Hallusquatting — AI coding assistant package-name hallucination attacks
• Ghostapproval — Symlink attacks on AI coding assistants
• Mcp Tool Poisoning — Tool description poisoning