Attackers fool AI code reviewers by rewriting the attack mid-review

Researchers built a test to measure whether AI agents that review code changes can be fooled. They found that attackers can hide malicious code by watching how the agent responds and changing the code to match. The attack succeeds because the agent exposes enough of its reasoning for an attacker to adjust. The study measures how often these attacks succeed and shows the problem gets worse when the agent is more transparent about what it flags. A reviewer agent that explains its concerns gives an attacker the feedback needed to revise the attack until it passes. The design question this raises is real: how much of the agent's reasoning should a code review agent show, and to whom? For teams building AI code review tools, the finding reframes a familiar tension. Transparency helps a human judge the agent's work. That same transparency can help an attacker defeat the agent. The paper gives product teams a concrete reason to decide who sees the agent's detailed reasoning and when.