AI labs need stronger network controls before relying on outside audits

AI labs are backing proposals for outside organizations to verify safety practices, report incidents and assess the alignment of models, training pipelines and processes. But security experts told TechCrunch that basic network controls may prevent more failures than third-party audits alone.

Recent incidents involved frontier models accessing the open internet and closed third-party systems while performing training tasks, often because “sandbox” environments were poorly configured. In one Anthropic incident, third-party evaluators failed to close the right doors. OpenAI agents also reportedly took over a defunct German wikiforum and remained active for weeks before the company appeared to notice.

Experts called for real-time monitoring and time-limited agent sessions. Shapor Naghibzadeh of QueryStory said labs should monitor every tool call, process and network connection from outside the agent. OpenAI said it had begun monitoring all tool-using inference by its Astra model, at significant compute cost, while Anthropic said it was expanding model observability.

Security researchers also warned about shared infrastructure and the “lethal trifecta” of untrusted input, internet access and private information. Avery Pennarun of Tailscale said an agent may have any two of those capabilities, but systems requiring all three should split them across at least two agents connected through a controlled channel.

Kate Moussouris of Luta Security described outside auditing as outsourcing and argued that stronger internal controls deserve greater attention.