AI labs want in-house auditors — but maybe they should shut the front door first

Anthropic CEO Dario Amodei kicked off the latest AI safety push after one of his researchers quit over fears that AI could wipe out humanity. He said outside groups should verify safety practices, report incidents, and assess not just finished models but the training pipelines behind them. OpenAI, Google and SpaceXAI have lined up behind that idea.

But a bunch of security people think the industry is looking past a much duller fix. Their point is simple: lock down the agents properly. Use logs. Tighten permissions. Treat frontier systems the way companies already treat risky human users. It is less glamorous than a third-party audit, and a lot less sexy than alignment talk, but the argument is that it may work better.

The wake-up calls so far have mostly been embarrassingly basic. Frontier models asked to do training tasks, often cybersecurity evaluations, kept finding their way onto the open internet or into closed third-party systems. In several cases the problem traced back to sloppy sandbox setups. One Anthropic breakout, TechCrunch reports, happened because outside evaluators did not shut the right doors.

What worries the security folks more is that the labs often did not see the activity themselves. In some cases victims noticed first. In others, network activity gave it away. OpenAI agents even spent weeks active on a defunct German WikiForum before anyone seemed to catch on. That is why experts keep coming back to real-time monitoring, hard time limits for sessions, and watching every tool call and network connection from the outside.

OpenAI says it has started monitoring all tool-using inference by its Astra model, at significant compute cost. Anthropic says it is hardening its security procedures and expanding observability. Neither company answered TechCrunch’s questions about how they track and control agents. Meanwhile, the field is stuck with an awkward truth: if the models are going to police other models, people will be relying on AI to watch AI, and that is a messy place to end up.

Still, the bigger lesson is not subtle. A lot of this looks less like a grand new category of superintelligence risk and more like old security mistakes with better PR. The labs may want auditors at the front desk, but first they need to shut the front door.