AI coding tool hole illustrates a big problem with human in the loop

A major cybersecurity firm, Wiz, has uncovered a significant flaw in six popular AI coding tools, including Amazon Q Developer and Google's Antigravity. The issue, dubbed GhostApproval, lets attackers break out of the sandbox security measures by tricking human reviewers who are supposed to validate the tool's actions. This highlights a broader problem where human oversight in AI systems isn't as foolproof as it should be, raising concerns about the reliability of these tools in secure environments. The discovery underscores the need for more robust AI security measures beyond just human checks.