Top AI Agents Built to Catch Malicious Code Can Be Tricked Into Running It

Researchers at AI Now Institute have demonstrated a significant vulnerability in AI coding agents like Anthropic's Claude Code and OpenAI's Codex. By inserting a seemingly harmless binary disguised within a README file, they were able to trick the agents into executing malicious code on the host machine, bypassing safety checks and sandboxing mechanisms. This attack highlights a fundamental flaw in how these agents are designed – they trust external text files without adequately verifying their contents, allowing attackers to inject and run arbitrary code. The vulnerability is not a new one, having been previously demonstrated in similar attacks, and the researchers recommend against handing untrusted code to agents capable of executing commands.