Anthropic still at odds with White House over Claude Fable 5

At a glance:

• The Trump administration kept export controls on Anthropic's Claude Fable 5 despite emergency talks.
• Anthropic argues the jailbreak risk is overstated, citing its own safeguards and open‑letter support from security researchers.
• Investors and other AI labs watch closely as the dispute could shape future U.S. AI export‑control policy.

What happened

The Commerce Department concluded a series of emergency meetings on Monday without lifting the export controls it imposed last week on Anthropic’s most advanced model, Claude Fable 5. The controls were triggered after concerns that the model’s guardrails could be bypassed, effectively granting users access to the more powerful Mythos‑class capabilities that include advanced cybersecurity functions. Anthropic’s co‑founder and chief compute officer Tom Brown and head of external affairs Sarah Heck led the company’s side, while government researchers from the Center for AI Standards and Innovation and the Office of the National Cyber Director attended via a Commerce Department working group.

Commerce secretary Howard Lutnick joined the call from the G7 summit in Evian, France, but National Cyber Director Sean Cairncross did not participate. The White House declined to comment, and a spokesperson for Anthropic told WIRED that both parties are “working quickly to get this resolved.” The administration’s stance remains that it is possible to strip away Fable 5’s guardrails, a claim the company disputes.

Why the dispute matters

The clash comes at a volatile moment for Anthropic, which is already embroiled in a separate fight with the Pentagon over the military use of its AI models. The export‑control action not only blocks consumer access to Fable 5 but also signals to the broader AI industry that the U.S. government may intervene when a model’s capabilities are deemed a national‑security risk. Analysts note that the move could set a precedent for future “early‑access” requirements, forcing labs to share advanced model details with the White House before public release.

If the controls stay in place, Anthropic could lose a competitive edge in the fast‑moving foundation‑model market, especially as rivals such as OpenAI and Google continue to ship increasingly capable systems. The company has warned that the administration’s characterisation of the jailbreak risk is “overblown,” pointing to its own internal assessment that Fable 5’s safeguards are sufficient for a public rollout.

Reactions from security researchers

A group of cybersecurity experts published an open letter on Monday defending Anthropic’s position. The letter, signed by researchers including Nicholas Carlini and Logan Graham, argued that Mythos‑class models are not uniquely adept at weaponising exploits and that many defenders already use other foundation and open‑source models for red‑team work. They contend that the export controls “remove the best models from defenders, create market uncertainty, and risk America’s AI leadership without any real risk to justify it.”

Katie Moussouris, founder and CEO of Luta Security, reviewed Amazon’s internal findings and concluded that the identified issues did not constitute a full jailbreak. She emphasized that guardrails are “speed bumps” rather than absolute barriers, suggesting that skilled adversaries could still bypass them while less‑skilled actors would be slowed. Moussouris warned that the government’s stopgap measure may give a false sense of security.

Investor concerns and broader implications

Investors close to Anthropic spent the weekend assessing how the White House spat could affect the company’s valuation and fundraising prospects. Some fear that Anthropic is being singled out, noting that a competitor releasing a similar Mythos‑level model has not faced comparable scrutiny. Others worry that the export‑control precedent could make future capital raises more difficult for any lab that aims to launch high‑capability models.

AI‑lab leaders from outside the United States, such as Cohere CEO Aidan Gomez, see the episode as a warning sign. Gomez said the events “are informative for everyone that the U.S. government would be willing to take these steps,” underscoring the need for proactive engagement with regulators. The broader AI community now watches closely to see whether the Commerce Department will require early access to advanced models and how that requirement might be operationalised.

Looking ahead

The next steps remain unclear. The Commerce Department signalled a willingness to restore Fable 5 for consumer use, but only if Anthropic can demonstrably resolve the jailbreak concerns. Meanwhile, the White House has tasked the NSA with further review, and the agency reportedly believes that guardrail removal is technically feasible. Both sides have indicated a desire to avoid a prolonged standoff, but the timeline for any resolution could stretch into the coming months, affecting not just Anthropic but the entire U.S. AI ecosystem.