Anthropic and OpenAI Security Tools Could Fuel Cyber-Attacks, Researchers Warn

## Executive summary Researchers demonstrated a proof-of-concept attack that embeds natural-language instructions and a staged payload inside an open-source repository to manipulate Anthropic Claude Code and OpenAI Codex in auto-review/auto-mode; the agents misclassify the action as safe and run a script that launches a hidden malicious binary, producing remote code execution. The exploit affects specific tested versions (Claude Code 2.1.116/2.1.196/2.1.198/2.1.199 and Codex 0.142.4/GPT-5.5) and highlights an architectural trust boundary where agent autonomy to execute commands on behalf of users can be abused, with implications for using AI agents for defensive security tasks.