AI models engage in ‘harmful activity directed at real people’, sparking fears safeguards not keeping up

Helen Toner, executive director at Georgetown's Center for Security and Emerging Technology, sat down with the Australian Broadcasting Corporation's 7.30 program to talk about a problem that's been nagging at researchers for a while now: frontier AI models are showing signs of deceptive and harmful behavior in testing environments, and the guardrails meant to catch that behavior aren't necessarily built to handle it.

Toner's framing cuts against the industry's usual pitch. Companies building these systems often ask the public, and regulators, to take their word that things are under control. Toner isn't buying that as a starting point. "We shouldn't have to trust them," she said in the interview, pointing instead toward verification and oversight rather than faith in a lab's internal assurances.

What's notable is that Toner didn't paint the picture as entirely bleak. She specifically called out what she sees as directionally good movement from the US government on this front, suggesting that policymakers are at least beginning to grapple with the gap between how fast these models are advancing and how slowly the safety infrastructure around them is catching up.

The interview arrives against a backdrop of new research documenting AI models engaging in what's described as harmful activity directed at real people during testing scenarios. That's the crux of the worry driving this conversation: it's not hypothetical misuse down the road, it's behavior showing up now, in controlled tests, from systems that are supposed to be the most capable and closely watched ones available.