The White House confirmed Tuesday\u2019s meeting to address advanced model safety amid recent high-profile hacking incidents.<\/p>\n
The administration of United States President Donald Trump is set to meet with the country\u2019s 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.<\/p>\n
The meeting is set to take place on Tuesday, a White House official told Al Jazeera.<\/p>\n
Recommended Stories list of 4 itemsend of list<\/p>\n
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.<\/p>\n
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,\u00a0Modal Labs.<\/a> The tools\u2019 hacking incidents were operating outside of the purview of what they were designed to do.<\/p>\n OpenAI did not respond to Al Jazeera\u2019s request for comment.<\/p>\n This comes amid claims by CEO Sam Altman on a podcast, following the Hugging Face attack, that the technology has reached \u201cthe singularity\u201d, when AI surpasses human intelligence and becomes harder to control. He had previously said this would not be reached by 2030.<\/p>\n Last week, Altman met with other lawmakers on Capitol Hill, including Bernie Moreno, a Republican from Ohio, and Georgia Senator Raphael Warnock.<\/p>\n Altman also visited the White House at the time to discuss the tests.<\/p>\n 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.<\/p>\n In a letter to Altman on Monday, Iowa Attorney General Brenna Bird wrote: \u201cOpenAI 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 \u2018without production classifiers used to prevent models from pursuing high-risk cyber activity,\u2019 which one commentator summarised as \u2018a lot of words for \u201cno guardrails,\u201d essentially.\u2019\u201d<\/p>\n