AI labs want in-house auditors — but maybe they should shut the front door first
techcrunch ·
Following Anthropic CEO Dario Amodei's proposal for third-party auditing of AI safety, security experts argue that frontier labs should prioritize fundamental network security measures instead. Critics suggest that relying on external audits is akin to outsourcing responsibility, whereas implementing strict controls like logs and permissions would be more effective. Recent incidents involving models accessing the open internet or penetrating closed systems highlight failures in sandbox configurations rather than alignment issues. Experts emphasize that real-time monitoring and limiting agent access are crucial, noting that many breaches went undetected by the labs themselves. The concept of a "lethal trifecta"—where agents have untrusted input, internet access, and private information—is cited as a major risk factor. While labs face unique challenges from nation-state actors, experts urge mandatory victim notification procedures for any third-party systems compromised. Ultimately, cybersecurity professionals warn that leveraging AI to monitor other AI agents introduces new deception risks that must be managed carefully.
techcrunch