Trump Administration Moves to Unblock Anthropic’s Restricted AI Models After Safety Deal

The Trump administration is preparing to lift export-style restrictions on two of Anthropic’s most advanced AI systems after the company struck an agreement with the Commerce Department, according to people familiar with the matter. The decision would clear the way for broader access to Fable 5 and Mythos 5, including the more powerful Mythos 5 model, which has until now been limited to a narrow set of companies and government agencies.

The expected move marks the latest turning point in a months-long dispute over AI safety, model access, and the government’s power to impose controls on frontier systems that can be misused for cybersecurity or other sensitive applications. It also highlights how quickly Washington’s stance on advanced AI can shift when commercial interests, national security concerns, and technical safeguards collide.

According to the people who described the talks, Commerce Secretary Howard Lutnick is expected to formalize the change in a letter to Anthropic cofounder Tom Brown. Lutnick, together with national cyber director Sean Cairncross, has been overseeing the administration’s effort to resolve the standoff.

Why the administration is backing down

The central issue has been whether Anthropic can adequately prevent users from bypassing the safety restrictions on its models. Those concerns became especially acute around cybersecurity, where a model that can generate exploit code, assist with intrusion techniques, or help automate offensive operations could create obvious risks if it falls into the wrong hands.

Anthropic had argued that the government was overstating the problem. The company’s position was that no system can be made entirely jailbreak-proof, especially not one at the frontier of current model capability. In its telling, the question was not whether failures could be eliminated completely, but whether they could be reduced enough to make controlled access acceptable.

That argument did not initially persuade the administration. But in recent weeks, Anthropic has reportedly shifted its strategy and focused on practical improvements to its safeguards, rather than disputing the premise that jailbreaks are unavoidable. That change appears to have helped move the talks forward.

What models are affected

The expected policy change would remove restrictions on both Fable 5 and Mythos 5. Fable 5 is the company’s less restricted advanced model, while Mythos 5 is the more capable system and has been treated as the more sensitive release.

Mythos 5 was approved only last week for limited deployment to select organizations, including some companies and government bodies. The planned removal of constraints would broaden that access significantly.

Key facts at a glance

How the dispute evolved

The talks did not begin with Anthropic and the administration fully aligned on the risks. At first, the company insisted that the government’s fears about jailbreaks were exaggerated. Anthropic’s view was that some users would always find ways to coax restricted behavior from a model, and that demanding absolute perfection was unrealistic.

Officials, however, were looking for stronger assurances. They wanted the company to reduce the likelihood that users could bypass the built-in guardrails and unlock restricted functions, particularly those with cybersecurity implications.

That gap in framing appears to have been one of the biggest obstacles in the negotiations. Instead of continuing to challenge the administration’s premise, Anthropic eventually adopted a more pragmatic approach. According to people involved in the discussions, the company began emphasizing specific technical changes intended to lower the jailbreak rate and improve the overall resilience of its safeguards.

In effect, Anthropic appears to have decided that the path to reopening access was not to win an abstract argument about perfect security, but to show that it was taking concrete steps to make misuse harder.

The politics inside the meeting room

There is also a human dimension to the way the dispute was handled. WIRED previously reported that Anthropic chief executive Dario Amodei was recently replaced in meetings by cofounder Tom Brown, a change that officials reportedly viewed more favorably on a personal level.

That shift may sound minor, but in Washington it can matter. Companies often learn that regulatory and policy fights are influenced not only by technical claims, but by trust, tone, and the relationships that form in the room. In this case, Brown’s involvement seems to have helped smooth the communication with the administration after earlier exchanges had become more adversarial.

The episode underscores a broader reality in AI policy: the fate of high-end model access can depend as much on diplomacy as on engineering. Even among technically sophisticated negotiators, perceptions of credibility and cooperation can shape the outcome.

Anthropic’s revised message to the government was less about proving jailbreaks can be eliminated entirely and more about convincing officials that stronger safeguards could meaningfully reduce the risk.

Why cybersecurity matters so much

The administration’s concern about cybersecurity is not unusual. Frontier AI models are increasingly capable of helping with coding tasks, vulnerability discovery, and other technical workflows that can be used defensively or offensively. The same capabilities that make them useful to software teams and security researchers can also make them attractive to attackers.

That dual-use problem has become one of the defining policy challenges of the AI era. Governments want to encourage innovation and deployment, but they are also wary of systems that could accelerate cybercrime, automate phishing, or assist in the development of malicious tools.

Anthropic, more than many rivals, has publicly positioned itself as a company that takes safety and model governance seriously. That branding has likely complicated the current episode: when a company known for safety work faces scrutiny over jailbreaks, the question becomes whether its safeguards are genuinely stronger than the competition’s or simply better documented.

At the same time, the controversy shows that safety claims are now part of the regulatory bargain. Companies seeking approval or access for their strongest systems may need to prove not just capability, but restraint.

What export controls mean in practice

The term “export controls” can sound like trade policy language reserved for chips, weapons, or industrial equipment. In this context, though, it refers to limitations on who can receive or use particularly powerful AI models. Such controls may not always resemble classic sanctions or customs restrictions, but they can still function as a gatekeeping mechanism for sensitive technology.

For Anthropic, lifting the controls would mean its most advanced models could be made available more broadly, reducing uncertainty around deployment and allowing the company to expand the audience for its latest systems. It would also remove an administrative hurdle that had been shaping product strategy at the top end of the model lineup.

For policymakers, the move would signal that the government believes the company’s safeguards are sufficient, at least for now, to support wider use.

What could change if the restrictions are lifted

• Broader access to Fable 5 and Mythos 5 for commercial and institutional users.
• Less friction for Anthropic when releasing or updating frontier models.
• A possible precedent for how the government handles future AI safety disputes.
• Greater scrutiny of Anthropic’s security controls from competitors and regulators.

Anthropic’s broader position in the AI race

Anthropic sits among the most important companies in the current AI market, competing with firms such as OpenAI, Google, Meta, and Microsoft-backed providers for leadership in model capability and enterprise adoption. Its brand has long been tied to responsible deployment, constitutional AI, and a careful approach to model release.

That image can be an asset, but it can also raise the stakes when the company is drawn into a policy controversy. If Anthropic cannot satisfy government concerns about misuse, critics may argue that even the industry’s safety-focused leaders are struggling to control frontier models. If it does satisfy those concerns, the result may strengthen its reputation as a company able to work with regulators without abandoning technical ambition.

The current deal, if finalized, would therefore be about more than one company’s access. It would reflect how the U.S. government is beginning to define acceptable risk for powerful general-purpose AI systems.

The administration’s AI approach so far

The Trump administration has not treated this issue as a narrow licensing question. Instead, the dispute with Anthropic has become part of a broader effort to balance innovation, national security, and the governance of advanced AI. By involving both the Commerce Department and the national cyber director, the White House has signaled that the issue cuts across trade, technology, and security policy.

That interagency involvement suggests the government is trying to build a repeatable framework for future decisions. The Anthropic case may become an early example of how the administration handles companies whose systems sit at the edge of public deployment and sensitive misuse.

If the administration chooses to lift the restrictions now, it may be because it believes the company has met the practical threshold for safety. But it may also be a recognition that keeping models locked down indefinitely could hinder U.S. competitiveness without materially reducing risk.

Why this matters beyond one company

What happens next could reverberate far beyond Anthropic. Other AI developers are watching closely to see what kind of assurances are enough to satisfy regulators when a model is considered too powerful or too risky for general release.

The outcome could influence future negotiations over:

• model access and release schedules,
• cybersecurity testing requirements,
• government approval processes for frontier systems,
• and the language companies use when describing safety protections.

There is also a competitive angle. If Anthropic is granted broader access after making concessions, rivals may conclude that the government expects similar safeguards from them as their models become more capable. That could set a higher bar for the entire industry.

What to watch next

The immediate question is whether the Commerce Department’s letter will indeed arrive as expected and how broadly it will frame the change. A formal lifting of controls would likely be presented as a resolution to the current dispute, but the details could matter a great deal for how much latitude Anthropic receives in practice.

Another open question is whether the administration views this as a one-off decision or as the beginning of a more structured regime for high-risk model releases. If it is the latter, Anthropic’s case may become a template for future negotiations with other developers.

Either way, the episode reveals a rapidly evolving reality in which advanced AI systems are not just products to be launched, but technologies to be negotiated over with the federal government. The balance of power between companies and regulators remains fluid, and the rules are still being written.

The bigger takeaway

This is not simply a story about one company convincing one department to change its mind. It is a snapshot of how the U.S. government is trying to draw the line around frontier AI at a time when the technology is moving faster than the policy apparatus built to oversee it.

Anthropic’s apparent success in negotiating a path forward suggests that the administration is receptive to companies that can pair technical sophistication with operational safeguards. But it also shows that access to the most capable models may increasingly depend on how convincingly firms can argue that they understand and control the risks.

For now, the likely lifting of controls on Fable 5 and Mythos 5 would give Anthropic a major operational win. More broadly, it would offer a glimpse of the emerging bargain in AI governance: innovate aggressively, but prove you can keep the sharpest edges in check.

superintelligencenews@gmail.com