Trump Admin Restores Anthropic Mythos 5 for 100+ Firms

Commerce Secretary Howard Lutnick laid out the reasoning in a letter to Anthropic's chief compute officer, Tom Brown, sent on Friday.

I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model.

The framing is deliberate. This is not a declaration that the model is safe for everyone. It is a narrow finding that a vetted set of organizations, under specific safeguards, can be trusted with it.

How the original ban worked, and why it happened

To understand the reversal, you have to understand what got pulled and why.

Mythos 5 is described as Anthropic's strongest cybersecurity model. Fable 5 is a version of Mythos 5 that was released to the public a couple of days before the ban. Fable 5 was said to carry more protections than its sibling, which is why it went out more widely in the first place.

Then both came down. The trigger was security research: the guardrails on the models were allegedly bypassed with relative ease. For a system built around cybersecurity tasks, easily-defeated safety controls are exactly the nightmare scenario, and the response was to remove both models entirely rather than risk misuse.

So the ban was never really about the technology being weak. It was about the safety layer being thin enough to peel back. That distinction is what makes a limited, supervised return possible: restrict who gets the model, wrap it in safeguards, and the risk calculus changes.

The critical-infrastructure angle

Anthropic confirmed the move publicly on Friday in a post on X, and the language points straight at why these particular organizations made the cut.

The company wrote that since June 12 it had been working closely with the US government to restore access to both Mythos 5 and Fable 5. It said the government had notified it that Mythos 5 can be redeployed to a set of US organizations that operate and defend critical infrastructure.

That phrase is the whole logic of the exception. Power grids, water systems, financial networks, and the agencies that protect them are under constant attack. A strong cybersecurity model is a defensive asset for exactly those defenders. Keeping it out of their hands while adversaries develop their own tools is its own kind of risk, and the administration appears to have weighed that trade-off and decided the defenders should be armed.

Fable 5 is still in limbo

One model is conspicuously missing from the good news. The directive did not address the release of Fable 5.

That is notable because Fable 5 was the one with more protections and the one that had already been released to the general public before the ban. Yet the reprieve covers only Mythos 5, and only for a controlled list of organizations.

Anthropic says it is not done. In its statement the company said it is restoring access for the approved organizations quickly, and that it is continuing to work with the government to expand access to Mythos 5 and to make Fable 5 available for general use again.

What to watch over the next 24 to 72 hours

The immediate phase is operational rather than political. Here is where attention should sit in the coming days.

• Speed of redeployment. Anthropic said it is moving quickly to restore access for the approved list. Watch for confirmation that the first agencies and companies are actually running Mythos 5 again, versus still waiting on paperwork.
• The approved list itself. Only the count, more than 100, is public so far. Any leak or disclosure of which agencies and companies are on it would reveal how broadly the administration defines critical infrastructure.
• Fable 5 signals. Because Fable 5 was left out, any follow-up letter or statement touching it would be the next real shift. A second directive is the thing to wait for.
• Researcher response. The guardrail bypass that started this is the unresolved core issue. Whether the safeguards Lutnick referenced actually hold up to fresh scrutiny will shape whether access keeps expanding or snaps shut again.

The bigger picture for AI policy

Strip away the specifics and this is a live test of how the US plans to govern frontier AI. The pattern here is ban first, then grant tiered exceptions based on who you are and what safeguards surround you, rather than a single rule applied to everyone.

It is a model of access by trust level. The most capable systems do not go fully public or stay fully locked; they flow to vetted operators under conditions. Whether that approach holds, expands, or collapses over the next few weeks will say a lot about how the next generation of powerful models reaches the market, and who gets to use them first.

Source: TechCrunch