Empty Is Not Clean: Five Fail-Open Bugs in an AI Agent

The article highlights critical vulnerabilities in an AI agent's security policy, revealing that five "fail-open" bugs allowed unauthorized access to sensitive areas like /etc. Despite a clear rule to deny access to such paths, a Bash call to /etc/shadow was inexplicably approved due to an overlooked bug. This oversight, replicated by multiple reviewers, underscores the risks of automated systems not having human oversight. These bugs emphasize the need for rigorous testing and human review in AI-driven security policies to prevent breaches.