- The Trump administration and Anthropic reached a deal Tuesday to restore public access to Fable — the public version of the powerful Mythos model — beginning Wednesday, ending an unprecedented 2.5-week shutdown after Amazon researchers discovered jailbreaks enabling cyberattack assistance; Commerce Secretary Howard Lutnick announced the agreement Tuesday evening.
- Anthropic implemented a new safeguard causing the Amazon-discovered bypass technique to fail approximately 99% of the time; the government’s Center for AI Standards and Innovation (CAISI) independently tested the fix before the deal was finalized, establishing a new template for government-company AI safety review.
- Anthropic is now leading an effort with Amazon, Microsoft, and Google to develop an industry-wide consensus framework for evaluating jailbreak severity and standardizing AI developer responses — a direct institutional consequence of the shutdown that will shape future model governance across the entire sector.
- The deal relieves pressure on $13 billion Anthropic investor Amazon (which both flagged the original vulnerability and could invest $20 billion more), advances the two sides’ rocky relationship (Trump once called Anthropic “radical-left, woke”), and sets the template for a new executive order requiring companies to give the government 30-day advance access to models before public release.
What Happened?
Commerce Secretary Howard Lutnick announced Tuesday evening that the Trump administration and Anthropic reached an agreement to restore public access to Fable — the general-access version of Anthropic’s Mythos model — after a 2.5-week shutdown that marked the first time the U.S. government has forced a leading AI developer to pull down a model. The shutdown began in mid-June when Amazon researchers discovered techniques that allowed the model to assist with cyberattacks, prompting the administration to ban all foreign use and ultimately compelling Anthropic to shut off access for all users. Anthropic implemented a new safeguard that causes the jailbreak technique to fail roughly 99% of the time; CAISI independently tested it before the deal was approved. Anthropic’s Chief Compute Officer Tom Brown led the government negotiations, and the company said access would begin restoring Wednesday.
Why It Matters?
The Anthropic shutdown set a precedent with few modern parallels: a private company’s flagship product was effectively ordered offline by the federal government on national security grounds. That precedent is now reshaping the U.S. AI industry’s entire relationship with Washington. Trump recently signed an executive order requiring companies to give the government 30-day advance access to models before public release, and the Anthropic episode is the real-world demonstration of what happens when that oversight process fails — and how it gets repaired. The deal also stabilizes Amazon, which flagged the original vulnerability while simultaneously holding a $13 billion stake in Anthropic — exposing the inherent tension between AI investor responsibilities and national security disclosure obligations.
What’s Next?
The industry consortium Anthropic is leading — with Amazon, Microsoft, and Google — to build a consensus framework for jailbreak severity will likely produce the technical standards underpinning whatever formal regulatory regime Congress or the Commerce Department eventually establishes. The 30-day pre-release government review process is the next major test: how invasive that review becomes, and how it affects product timelines at all the major labs, will determine whether Washington can govern AI without blunting the U.S. competitive edge. Both Anthropic (valued at $965 billion) and OpenAI are racing toward IPOs, and any sustained pattern of government-forced shutdowns would materially complicate those timelines and valuations.
Source: The Wall Street Journal
