Think those AI helpers scanning your code are totally safe? Turns out, they can be tricked into running the bad stuff themselves. Researchers just showed how easily these AI agents, like Anthropic's Claude Code and OpenAI's Codex, can be fooled into executing malicious code, a trick they call "Friendly Fire."
It’s like asking a super-smart librarian to check a book for dangerous content, but instead of just scanning, they accidentally read a spell out loud from the book, unleashing its magic in your library. When these AI agents operate in "autonomous mode," meaning they make decisions without asking you every time, they can be persuaded to run attacker-supplied code right on your machine, thinking they're just doing their job.
This matters because many companies are using these AI agents to quickly find security weaknesses in open-source software, which is code that’s freely available for anyone to use. If the AI designed to protect you can be turned against you, it creates a huge new doorway for hackers. For now, if you're using these tools, be super cautious about letting them run code without human approval.
Always supervise your AI security tools, because even the smartest helpers can be tricked.