# Claude Mythos 5 [[Anthropic]]'s most capable AI model and the first to sit in the "Mythos class" — a tier above Claude Opus in the company's model hierarchy. Mythos 5 is the same underlying model as [[Claude Fable 5]] but with safety guardrails lifted in specific capability domains, particularly cybersecurity and molecular biology. It is not generally available; access is restricted to vetted cyberdefenders, critical infrastructure providers, and U.S. government agencies through a program called Project Glasswing. Mythos 5 matters for the KB as the model that triggered the first AI-specific export control directive in U.S. history — when the government ordered [[Anthropic]] to suspend all access by foreign nationals in June 2026, citing a jailbreak vulnerability. The episode crystallized the tension between frontier AI capability and national security control that the KB tracks across the [[Dual-Use AI]] landscape. ## Capabilities Mythos 5's defining feature is cybersecurity: it can find software vulnerabilities better than all but the most skilled human security researchers. On Anthropic's internal cybersecurity evaluations, Mythos 5 scores 78.0% — nearly double Opus 4.8's 40.0%. In blinded comparisons on molecular biology hypotheses, Anthropic's scientists preferred Mythos 5's output roughly 80% of the time over Opus-class models. Both Mythos 5 and [[Claude Fable 5]] share the same technical specifications: a 1M token context window, up to 128K output tokens per request, and pricing at $10/$50 per million input/output tokens (2x Opus 4.8). ## Project Glasswing Anthropic launched Project Glasswing in early April 2026 with Claude Mythos Preview — a precursor to Mythos 5 — deployed to roughly 50 initial partners to scan codebases for vulnerabilities. By June 2, 2026, the program had expanded to approximately 150 organizations across 15+ countries. Early results were striking: partners found more than 10,000 high- or critical-severity security flaws. Anthropic's own scan of 1,000+ open-source projects flagged 23,019 potential vulnerabilities, with over 90% of independently reviewed high/critical findings confirmed as valid. The program's success created its own political problem: a model that can find vulnerabilities in every major operating system and web browser is also a model that adversaries would desperately want access to — the logic that drove the June 2026 [[Anthropic Model Export Directive]]. ## The Export Directive and Partial Restoration On June 12, 2026, the U.S. government issued an export control directive ordering Anthropic to suspend all Mythos 5 and [[Claude Fable 5]] access for foreign nationals — including foreign-national Anthropic employees. The stated rationale was the discovery of a method to bypass Fable 5's safety guardrails (a jailbreak), which raised fears about unrestricted Mythos-class capabilities reaching adversarial actors. On June 27, 2026, Commerce Secretary [[Howard Lutnick]] signed a partial restoration: approximately 100 trusted U.S. companies and federal agencies regained access to Mythos 5 specifically, after Anthropic executives — including co-founder Tom Brown — negotiated additional safeguards with senior administration officials. [[Claude Fable 5]] remained suspended as of that date. ## Why It Matters Mythos 5 is the KB's clearest example of frontier AI capability outrunning the regulatory apparatus designed to contain it. The model is simultaneously too powerful to share freely (hence the export directive) and too powerful to restrict entirely (hence the partial restoration for critical infrastructure). The government's response — an ad hoc executive action rather than legislation — reflects the absence of any statutory framework for AI export control, a gap the [[Dual-Use AI]] debate has not yet filled. The episode also extended Anthropic's collision course with the Trump administration, which had already banned the company from government contracts in February 2026 over its refusal to remove military-use restrictions (the [[Anthropic Government Ban]]). The Fable/Mythos directive was imposed on different grounds — export control rather than contract compliance — but reinforced the pattern of the administration using executive power to control which AI capabilities are available and to whom. ## See Also - [[Claude Fable 5]] — the publicly available, safety-guardrailed version of the same model - [[Anthropic]] — the company behind both models - [[Anthropic Model Export Directive]] — the June 2026 foreign-national access ban - [[Anthropic Government Ban]] — the earlier February 2026 Pentagon contract ban - [[Dual-Use AI]] — the framework for AI that serves both civilian and military purposes - [[Project Maven]] — the Pentagon AI program Anthropic was removed from - [[Howard Lutnick]] — the Commerce Secretary who signed the partial restoration