Trump’s plan to test AI models has a problem – the US security forces were fired by the DOGE

When the masked examples were revealed, Nguyen warned that security measures would depend on whether the AI ​​industry was transparent and considered the project a “true partnership.”

“Underneath the problem of definition lies a problem of perception,” Nguyen wrote. “The government can’t analyze what it can’t see, and the potential of the frontier is only seen by the labs that create it.”

Ferren noted that “the window for developing proper cyber defenses for new AI models may be closing quickly,” and that even the most well-designed government program may struggle to accurately predict the borderline models in the short term. “Even if done well, pre-deployment testing has limitations,” said Ferren, noting that Google’s threat intelligence team has found government affiliates using borderline models to carry out cyberattacks and “investigators have found. displayed that the horrors of the Mythos can be recreated with an open-source heavy machine. “

Therefore, even if AI can volunteer to be tested, they may be interested in money to get a rubber stamp, rather than working with the government to test the known potential of their limits in full.

“It would be difficult to create a model that would not be harmful to abuse but still be commercially compelling,” Ferren said.

He added that the EO “may provide short-term cybersecurity benefits,” but the “long-term consequences” remain “unclear.”

Nguyen said the EO should take steps to establish “cyber monitoring, voluntary monitoring, and threat analysis” that “the national security community will need for many years” to “continue to evaluate systems that are possible rather than deterministic, independent rather than controlled, and which are subject to change with every change.”

But security testing must evolve as quickly as technology does, Nguyen said, otherwise we risk testing “yesterday’s threats”.

That’s why, at the core, the process will depend on honest exchanges between people with deep expertise and secret knowledge of national security. It’s the only way to ensure that the US is focusing its efforts on protecting people from credible and consequential AI threats, rather than simply providing “effective guarantees,” Nguyen wrote.