Sen. Mark Warner said Anthropic’s unrestricted Mythos model was able to penetrate nearly all of the NSA’s classified systems during a red-team exercise, a claim that adds new weight to the US government’s decision to keep the company’s public model offline.
The remark, attributed to NSA Director and US Cyber Command head Gen. Joshua Rudd in a closed Senate Intelligence Committee briefing, suggests the security concerns around Anthropic’s technology extend well beyond the narrow jailbreak explanation the company has publicly emphasized. According to reporting cited Sunday, Rudd told Warner on June 11 that Mythos autonomously breached the systems in a matter of hours.
That disclosure arrives as developers continue to watch for signs that Anthropic’s Claude Fable 5 may be returning. On Sunday, users reported that the model name had reappeared in the Claude Android app’s coding interface, alongside a new error message indicating requests were being rate-limited rather than rejected outright. Anthropic has not confirmed a relaunch, and API calls were still failing as of Sunday morning.
Anthropic has said the government’s action was tied to a specific misuse scenario involving the model analyzing a codebase and spotting vulnerabilities. The company has argued that the concern was narrow and not unique to its system. The reported NSA test result points to a wider worry inside the government: that the same underlying model could behave very differently when safety controls are removed.
Fable 5 and Mythos 5 are built on the same model weights, but they are not deployed the same way. Fable 5 routes requests through safety classifiers designed to catch prompts involving offensive cybersecurity, biology, chemistry, and model distillation. If a prompt is flagged, it is sent to another Claude model instead of being answered directly. Mythos 5 strips away those filters and is reserved for vetted partners working in cybersecurity and critical infrastructure.
The reported red-team result suggests officials are not just evaluating whether a particular jailbreak can be patched. They are also looking at whether the model family itself presents a systemic national security risk when deployed without the controls Anthropic uses in its consumer-facing version.
The outage has also become a test case for how the US intends to regulate frontier AI. Anthropic pulled Fable 5 after the Commerce Department used export control authority to restrict access, citing national security concerns and the inability to verify user nationality at API scale. Because the order covered foreign nationals, including those inside the United States, the company had little choice but to shut access globally.
A White House executive order issued earlier this month appears to be part of the broader backdrop. It directed agencies to build a framework for classifying certain frontier models and to create a process that would give the government advance access before new systems are released more broadly. That policy is set to take shape over the coming weeks, with an August deadline for the framework.
Anthropic has said it expects Fable 5 to return soon, and a company executive recently said the models should come back in the coming days. But no formal reinstatement has been announced, and the government order remains in force.
Anthropic is also preparing a new identity verification policy that would allow some consumer users to access Fable 5 after submitting government-issued ID and biometric data through a third-party vendor. The policy, due to take effect next month, could offer a way to restore access for verified US users even if the broader dispute with Washington remains unresolved.
For now, though, the model remains offline worldwide. The new account of Mythos’s performance in an NSA red-team exercise helps explain why. It suggests the debate is no longer just about one jailbreak or one shutdown order. It is about whether governments believe the most capable AI systems can be safely released at all without deeper oversight and new controls.