White House to meet AI firms on advanced model safety
The administration of United States President Donald Trump is set to meet with the country’s top artificial intelligence companies to discuss allowing the federal government
The administration of United States President Donald Trump is set to meet with the country’s top artificial intelligence companies to discuss allowing the federal government more safety tests for its most advanced models amid two high-profile hackings confirmed by Anthropic and OpenAI. The meeting is set to take place on Tuesday, a White House official told Al Jazeera. The meeting is part of a June 2 executive order that called for a new voluntary framework that would allow the federal government to have early access to advanced models for up to 30 days before public release, it said. The meeting is underscored by developments last week when OpenAI tools were responsible for a breach at a tech company, Hugging Face, and at the client of another, Modal Labs. The tools’ hacking incidents were operating outside of the purview of what they were designed to do.
OpenAI did not respond to Al Jazeera’s request for comment. This comes amid claims by CEO Sam Altman on a podcast, following the Hugging Face attack, that the technology has reached “the singularity”, when AI surpasses human intelligence and becomes harder to control. He had previously said this would not be reached by 2030. Last week, Altman met with other lawmakers on Capitol Hill, including Bernie Moreno, a Republican from Ohio, and Georgia Senator Raphael Warnock. Altman also visited the White House at the time to discuss the tests. On Monday, 15 state attorneys general, all Republicans, including those from Florida, Iowa and Kansas, called on OpenAI to maintain documents on how its AI system escaped containment. In a letter to Altman on Monday, Iowa Attorney General Brenna Bird wrote: “OpenAI conducted that testing in what should have been an isolated environment, with no possibility that the agent could access the Internet.
Within that purportedly isolated environment, OpenAI allowed the agent to operate ‘without production classifiers used to prevent models from pursuing high-risk cyber activity,’ which one commentator summarised as ‘a lot of words for “no guardrails,” essentially.’” Anthropic also announced last week that it had identified three separate incidents in which its model, Claude, was able to reach the internet and gain unauthorised access to an undisclosed set of organisations. The Dario Amodei-led AI giant said that the hacking incidents occurred in a “third-party evaluation environment”, according to a release from the company. Anthropic has been in the crosshairs of the White House over its efforts to restrict the government’s use of its technology for autonomous weapons and mass surveillance. Anthropic did not respond to Al Jazeera’s request for comment. Google is expected to be in attendance, according to reporting by the tech outlet The Information.
The BriefWire