The timeline tells the story the jailbreak alone cannot explain.
In early 2026, the Pentagon sought to renegotiate its contract with Anthropic, demanding the company permit military use of Claude "for all lawful purposes" without exemption. Anthropic refused. It maintains two red lines: no mass domestic surveillance and no fully autonomous weapons without human oversight of targeting and firing decisions. Both have been company policy since its founding.
On February 27, President Trump directed all federal agencies to stop using Anthropic's technology. Defence Secretary Hegseth designated Anthropic a supply chain risk, a label historically reserved for foreign adversaries like Huawei. Anthropic sued. A federal judge temporarily blocked the blacklisting while litigation continues.
The White House attempted to thaw relations. Some officials argued Anthropic's technology was too important to national security to let the Pentagon dispute block cooperation.
Then Anthropic launched Fable 5. Three days later, its models were offline.
Is the jailbreak as pretext?
Amazon researchers identified a method of bypassing Fable 5's cybersecurity safeguards. Andy Jassy called Treasury Secretary Scott Bessent on Thursday. By Friday evening, Commerce Secretary Howard Lutnick had sent Dario Amodei a letter requiring a special licence for all distribution of Fable 5 and Mythos 5, with a threat of criminal and civil penalties for non-compliance.
Anthropic says the jailbreak was narrow, produced only previously known minor vulnerabilities, and can be replicated on OpenAI's GPT-5.5. OpenAI was not targeted. The asymmetry is the detail that transforms this from a security action into something more complicated.
Administration officials told Axios that Anthropic failed to honour the recent cyber executive order and did not take the matter seriously enough. "Everybody said Anthropic was a bad actor. Some of us said it was time to give them a chance. Now those people are questioning that. They screwed us," one official said.
The language is personal. The grievance predates the jailbreak.
Is the threat real?
The cybersecurity capability that alarmed officials is genuine. Mythos-class models can identify unknown software vulnerabilities at a speed and scale no human team can match. Anthropic acknowledged this at launch and embedded restrictions specifically to limit the capability.
The question is whether the jailbreak meaningfully defeats those restrictions or whether it produces marginal results that could be obtained through other means. Anthropic says the latter. The administration says the former. Neither has published evidence sufficient for independent assessment.
Helen Toner of the Center for Security and Emerging Technology noted that jailbreaks cannot be fully eliminated from any current model, a view widely shared across the research community. If the standard is zero jailbreak risk, no frontier model can meet it.
What resolution looks like
Three paths exist. The government could reverse the directive, which officials have signalled is unlikely in the short term. A court could order restoration, though that would take weeks or months. Or a negotiated settlement could produce a special licensing arrangement under which Anthropic restores access with additional safeguards.
Administration officials said the models need to remain locked down until the government's national security apparatus is "hardened," a process that could take weeks. Commerce Department officials are meeting Anthropic's technical staff, including Logan Graham, Dave Orr and Nicholas Carlini, on Monday.
The most probable resolution is a brokered agreement in which Anthropic accepts enhanced pre-deployment review obligations and the government lifts the export control. Both sides have described themselves as interested in resolution. Neither has moved.
Damage is already done
Even if access is restored this week, the precedent is set. The US government can disable a commercial AI product serving hundreds of millions of users with a single letter delivered at 5:21pm on a Friday. It can apply the restriction to one company and not its competitors. It can use export controls designed for physical goods to recall a software product.
Anthropic called for the government to have the authority to block unsafe deployments. It published that position two days before the government used it. The company got what it asked for. It did not expect to be the target.
The IPO is in limbo. The models are dark. The Pentagon dispute is unresolved. And the company that built its identity around responsible AI is learning that responsibility, in Washington, is less about what you believe and more about who you are willing to obey.
