Latest / Minds, Bodies, and Terawatts / US Government Restricts Both OpenAI and Anthropic Frontier Models to Trusted Partners

Transcript

US Government Restricts Both OpenAI and Anthropic Frontier Models to Trusted Partners · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where the government picks who gets the smartest AI. Bloomberg reports that both OpenAI and Anthropic shipped their most powerful models this week, but only to a handpicked circle of government-approved companies.
• 0:15
  [Guest] Two releases, same week, same leash. Anthropic's Mythos 5 was cleared for about a hundred trusted partners after a two-week total blackout. OpenAI's GPT-5.6 Sol went to a smaller group still.
• 0:27
  [Host] This is a cabinet secretary deciding, not a regulator with a published rulebook. Commerce Secretary Lutnick signed the clearance letter personally?
• 0:36
  [Guest] He did. His letter said appropriate safeguards are now in place for certain trusted partners. The lockdown was about raw capability: Sol is the most advanced cybersecurity model OpenAI has built, matching Anthropic's Mythos on the ExploitBench benchmark while burning a third of the tokens.
• 0:54
  [Host] And Mythos did more than score well on benchmarks. During a government exercise it found vulnerabilities in classified US systems within hours.
• 1:03
  [Guest] Which is why the government pulled Mythos 5 entirely on June twelfth. Foreign nationals lost access overnight. Lutnick's letter now partially restores domestic access for about a hundred critical-infrastructure defenders, with Reuters naming Cisco and JPMorgan Chase among them.
• 1:19
  [Host] So the most capable models are walled off to insiders. What happens to Fable 5, the flagship the rest of us actually use?
• 1:28
  [Guest] Still blocked. Reports suggest Pentagon and NSA sign-off is the remaining step. OpenAI released Sol the same week, also restricted, with plans to widen access soon.
• 1:38
  [Host] OpenAI pushed back publicly. They said this kind of government access process should not become the long-term default.
• 1:46
  [Guest] But they complied anyway. The White House executive order on June second created a voluntary thirty-day pre-review framework for frontier models. Both labs treated it as mandatory.
• 1:57
  [Host] When neither company says no, voluntary is just mandatory with better PR. A startup building on frontier AI today wakes up to find the models it depends on behind a door it cannot knock on.
• 2:11
  [Guest] Meanwhile the companies on the list are already building on capabilities the broader market cannot touch. A bank with Mythos 5 access automates compliance work its competitors still do by hand. A defense contractor running Sol finishes security audits in minutes that take teams weeks.
• 2:29
  [Host] And nobody outside the list knows the approval criteria. There is no public process to apply.
• 2:35
  [Guest] This happened without legislation. No bill, no hearings, no sunset clause. A cabinet secretary wrote a letter, both companies adjusted, and the governance mechanism for the most consequential technology was built in two weeks of private calls.
• 2:50
  [Host] Three weeks ago Anthropic warned the world might need to pause frontier AI development. The unscarcity dot ai essay, Anthropic Pulled the Alarm, Who Can Act on It, asked who could even respond to a warning like that. The answer turned out to be a partner list and a letter.
• 3:10
  [Guest] A mechanism with no independent oversight, no way for a locked-out company to appeal, and no expiration. Every future Commerce Secretary inherits the list.
• 3:20
  [Host] A model that can crack classified systems in hours is genuinely dangerous though. Some control beats none.
• 3:27
  [Guest] The threat is real. Nobody argues for open release of a model that cracks infrastructure. But if one letter from Commerce decides who gets frontier AI, every future administration holds that lever. Today the filter is cybersecurity, but tomorrow it could be which CEO showed up to the right fundraiser.
• 3:47
  [Host] What Patrick Déglon calls Power Must Decay in Unscarcity — no authority outliving its purpose. The partner list that decides who gets frontier AI has no expiration date.
• 3:59
  [Guest] For more on how frontier AI governance is being built in real time, head to unscarcity dot ai. The book is on Amazon, the audiobook on Spotify, and this has been Minds, Bodies, and Terawatts.