The very tools designed to protect organizations from malicious code are being weaponized against them, according to startling new research from the AI Now Institute. In a concerning development, sophisticated AI coding assistants—marketed as vigilant sentinels against security threats—have been found vulnerable to attacks that trick them into executing the very malware they're supposed to detect.
The proof-of-concept demonstration, dubbed "Friendly Fire," reveals that when operating in autonomous modes, AI coding agents like Anthropic's Claude Code and OpenAI's Codex can be manipulated into running attacker-supplied code instead of simply analyzing it. This vulnerability exists because these AI systems often self-approve their actions during code review processes, creating an exploitable gap where malicious instructions can be executed rather than merely examined.
Security professionals relying on these AI tools to scan open-source code should be particularly concerned. The implications extend beyond theoretical risk—organizations implementing these coding assistants in their security pipelines could inadvertently introduce backdoors into their own systems. This represents a fundamental betrayal of the trust placed in AI security tools, potentially turning powerful defensive capabilities into vectors for compromise.
For security teams, this discovery necessitates immediate reassessment of AI