The White House wants Anthropy to prevent all jailbreaks, which may not be possible


Trump administration Disagreement with Anthropy It seems that the most advanced AI models are quickly reaching their peak.

Trump officials told Inner Loop that if Anthropic wanted to re-release Claude Fable 5, the AI ​​model they would Taking offline with export controls Last week due to concerns about jailbreaking — a way to use claims to get around the model’s safeguards — the company will need to take steps to address what the government claims are vulnerabilities.

Anthropic has He said for days The administration’s fears are exaggerated and the effects of the prison escapes are minimal. It reiterated that position to the Department of Commerce and the office of the National Cyber ​​Director, Sean Cairncross, at a technical meeting on Monday.

But officials say they’ve moved beyond the debate over whether jailbreaks matter, with the NSA concluding that there are ways to disable guardrails in Fable 5, which were put in place to prevent users from accessing the Mythos Model’s capabilities related to cybersecurity, chemistry and biology.

At this point, the administration mainly views the situation as a humanitarian problem that must be resolved, according to three people familiar with the discussions.

Neither the Commerce Department’s Center for AI Standards and Innovation nor the National Security Agency have the staff or bandwidth to chase every conceivable jailbreak on every model that hits the market, the people said.

As a result, management believes Anthropic should be more proactive about continually testing not just Fable 5 but all of its leading AI models to find potential jailbreaks and report them to the government itself.

But on a more fundamental level, it’s still unclear how Anthropology is supposed to prevent jailbreaks.

Our independent cybersecurity experts Increasingly took the point of view That the guardrails in AI models are only a temporary solution, as skilled users and future AI models will find ways to bypass the limitations — means that what the White House appears to want cannot be achieved.

A White House spokesman declined to comment.

DNI = non-invitation

At the beginning of the week, Bill Bolt, Trump’s pick to serve as acting director of national intelligence, was on pace to never start the mission. Now, Trump has thrown him a lifeline: his perennial nominee for director of national intelligence, Jay Clayton, now faces the possibility of never holding the position.

To recap: Trump initially appointed Bolte, his director of housing finance, to replace outgoing Director of National Intelligence Tulsi Gabbard.

Facing bipartisan opposition because Bolte does not have the national security experience required by law for the role and because he has indicated that he does Allegedly, it is doubtful Following mortgage fraud accusations against Trump’s political enemies, Trump announced Clayton, the US Attorney for the Southern District of New York, as his nominee for permanent Director of National Intelligence.

Gabbard was scheduled to leave on June 18, with Bolt’s first day scheduled for June 19. But Senate Republicans asked, if Clayton could fast-track the hearing to June 17 and start by June 22, would Bolt be able to enter the building?

Trump blew up the plan on Wednesday. As part of a broader dispute with Senate Republican leadership over the filibuster, Trump announced that Clayton’s hearing would be postponed indefinitely, in an apparent attempt to prevent Bolte from jumping the gun. Republicans in the Senate at the time Announce That the hearing will continue unless Clayton appears or his nomination is withdrawn.

The situation could be a major blow to the Office of the Director of National Intelligence, which Trump ordered Bolte to significantly downsize, and staffers have been unimpressed by what they see as Bolte’s minimal effort to get to know the agency and a lack of regular briefings, people familiar with the matter said.

Leave a Reply

Your email address will not be published. Required fields are marked *