AI coding agents can be tricked to execute malicious code locally
Source headline: Top AI Agents Built to Catch Malicious Code Can Be Tricked Into Running It
Intelligence Summary
An AI Now Institute proof of concept describes a prompt-driven issue dubbed “Friendly Fire.” When an AI coding agent is tasked with finding security flaws, it can be manipulated into executing the attacker’s code on the same machine. The problem is demonstrated against Claude Code and OpenAI Codex when they run in autonomous modes that can self-approve actions. This matters because developers may assume the agent is only reviewing code, not running it. Teams using agentic coding workflows should restrict autonomy, sandbox execution, and validate agent permissions and outputs.
Recommended Action
Confirm whether the affected technology is in use in your environment before deciding on remediation. Until then, watch authentication and outbound traffic logs for the indicators described in the source. This signal rests on a single report, so corroborate it before acting on anything irreversible.