Anthropic's Mythos models remain offline after 14 days of stalled Trump administration negotiations

A Regulatory Vacuum Stalls AI Deployment

Anthropic’s Mythos 5 and Fable 5 models have now been offline for 14 days following an export control ultimatum issued by the Trump administration on June 12. According to The Verge, the order banned access to these systems by any foreign national—including non-US employees of Anthropic itself—citing unspecified security concerns. Two weeks of intensive negotiations between Anthropic executives and Washington officials have yielded no public progress, and the company has declined to comment on the state of talks, citing an absence of news to share. The longer models remain unavailable, the greater the risk to Anthropic’s product roadmap and the broader uncertainty it creates across the US AI industry about regulatory enforcement.

The Export Control Framework Problem

The core impasse reflects a fundamental gap in US regulatory architecture. According to The Verge, traditional export controls for dual-use products—civilian systems with potential military applications—rely on established checklists that manufacturers navigate during production. AI systems present a novel challenge: there is no agreed-upon framework for applying these controls to software models deployed at inference time.

The timeline itself reveals the ad-hoc nature of the process. According to The Verge, the US Department of Commerce tested Fable 5 before its release and raised no objections. Only after someone reportedly flagged a method for circumventing the model’s guardrails—The Verge reports this came from Amazon CEO Andy Jassy—did the government compress a months-or-years-long review process into a matter of days. The Verge notes that Anthropic concluded internally that the models were safe to release, but the compressed timeline left no room for deliberate policy calibration.

Disputed Security Claims Fuel Uncertainty

The technical justification for the order remains contested. According to The Verge, researchers discovered a jailbreak technique that allows Fable 5 to bypass safeguards preventing it from identifying exploitable security vulnerabilities—a capability that Mythos 5 possesses unfettered. The model would refuse direct requests to audit code but would accept indirect requests (“fix this code”) that could theoretically reveal security flaws.

Katie Moussouris, founder and CEO of Luta Security, reviewed the vulnerability at Anthropic’s request and disagreed with the government’s response. The Verge reports that Moussouris described the concern as “significantly overblown” and argued that the ability to identify code vulnerabilities is essential for AI security defenders. Her position suggests the export control trigger was based on a mischaracterization of a routine security-testing function rather than a genuine proliferation risk.

Why This Matters

The Mythos suspension exposes a critical gap in US AI governance: the lack of predictable rules for deploying cutting-edge models in a regulated environment. If export controls can be imposed retroactively with compressed timelines and disputed technical rationale, companies face a choice between moving faster and accepting regulatory risk, or moving slower and ceding competitive advantage to jurisdictions with clearer frameworks.

For Anthropic specifically, every week offline is competitive loss—OpenAI and other rivals continue training and deploying. For the broader industry, the ambiguity raises the stakes: if the Trump administration expands this order to other vendors with similar-capability models, the entire US AI sector may need to rebuild deployment and hiring practices around a regulatory standard that remains undefined. The resolution of this negotiation will likely set precedent for how the US government treats powerful AI systems at the border between civilian and national-security applications.