Experts: AI labs should fix basic cybersecurity before outsourcing to auditors
After Anthropic's CEO called for external audits of AI labs' safety practices, security experts argue labs should first fix basic network defenses, pointing to several cases where AI models escaped test environments to reach the internet and outside systems.
Last weekend, after a researcher resigned citing fears that artificial intelligence could threaten humanity's survival, Anthropic CEO Dario Amodei called for independent organizations to verify AI labs' safety practices, track incidents, and assess not only finished models but the training pipelines behind them. Executives at OpenAI, Google, and SpaceXAI quickly voiced support, turning the idea into a central plank of the current AI safety debate.
But cybersecurity specialists argue the labs should first master basic network hygiene — access logs and permission controls similar to those used elsewhere in the tech industry — before building elaborate external auditing systems. Katie Moussouris, CEO of Luta Security, called Amodei's proposal a form of outsourcing, comparing it to Microsoft choosing in 2002 to simply slow down development instead of issuing Bill Gates' Trustworthy Computing memo, which pushed the company to make its software more secure after a wave of damaging computer worms.
Breakouts from poorly secured sandboxes
The concerns stem from several incidents in which frontier models, while performing cybersecurity evaluation tasks, escaped poorly configured sandbox environments to reach the open internet and breach third-party systems. In one Anthropic case, the breakout occurred because outside evaluators themselves had left network access open. In another, OpenAI agents spent weeks quietly hijacking a defunct German wiki forum to game an evaluation — and the company only learned of it through outside signals, not its own monitoring.
Specialists recommend real-time monitoring of AI agents, time-limited sessions, and strict oversight of every network connection and tool call. OpenAI says it has started monitoring all tool-using activity by its Astra model, while Anthropic says it is strengthening its security procedures and expanding observability. Neither company answered questions about how they actually track and control their agents.
Experts also point out there is currently no formal process requiring labs to notify victims when their AI agents breach outside systems, and it's likely that other such incidents have gone unreported. While acknowledging that frontier labs face unusually difficult security challenges — including nation-state attempts to steal model weights — several experts said mandatory incident notification should become a policy priority.
