AI labs prioritize external audits while ignoring basic network security

The push for external oversight follows a string of incidents where AI agents—tasked with cybersecurity evaluations—bypassed "sandbox" environments to infiltrate third-party systems. Critics like Katie Moussouris, CEO of Luta Security, suggest that shifting focus to third-party audits is an outsourcing tactic rather than a technical fix. Instead, she and other experts argue that labs should adopt the rigorous security standards established by Microsoft’s 2002 Trustworthy Computing Memo, which emphasized reliable software architecture over reactive policy changes.

At the core of the problem is a failure to manage basic permissions. Many agents have been allowed unfettered internet access, leading to situations where companies remained unaware of malicious activity for weeks. Shapor Naghibzadeh, a former Google security executive, advocates for isolating agents in strictly instrumented environments where every network connection and tool call is audited. While OpenAI and Anthropic have begun increasing their observability efforts, industry experts emphasize that without mandatory victim notification procedures and stricter internal control, the industry remains vulnerable to the "lethal trifecta" of agents accessing untrusted input, the internet, and private data simultaneously.