Mythos 5 Returns After 14-Day Jailbreak Ban: AI Safety Vetting Sets New Precedent

From a research and safety engineering perspective, the discovery of a jailbreak exploit in Fable 5—prompting the Commerce Department to freeze two cutting-edge models—is a red flag for the entire AI field. The government’s subsequent decision to allow Mythos 5 only for “cyber defenders” suggests that direct application in defensive contexts may be seen as a safer use case, while general-purpose deployment remains too risky, highlighting an evolving safety framework for frontier AI.

On June 26, 2026, the Trump administration partially reversed a two-week suspension on Anthropic’s most advanced AI model, Claude Mythos 5, authorizing access for a “small group of cyber defenders and infrastructure providers.” The move came after an abrupt June 12 export-control directive halted both Mythos 5 and its general-use sibling model Fable 5, following the government’s identification of a jailbreak vulnerability in Fable 5. While the Commerce Department’s letter signed by Secretary Howard Lutnick allows carefully vetted entities to once again tap Mythos 5, Fable 5 remains locked down, and no timeline or criteria for broader release were disclosed. Simultaneously, OpenAI revealed that its new GPT-5.6 model would also be restricted to an administration-approved user group, signaling a coordinated turn toward government gatekeeping of frontier AI.

“

While the Commerce Department’s letter signed by Secretary Howard Lutnick allows carefully vetted entities to once again tap Mythos 5, Fable 5 remains locked down, and no timeline or criteria for broader release were disclosed.

The Trump administration’s use of export-control authority to throttle access to private-sector AI marks a pivotal expansion of the national security state into software deployment. By classifying advanced models as dual-use technology subject to trade restrictions, the Commerce Department is treating today’s AI much as the Clinton administration treated strong cryptography in the 1990s. The immediate trigger—a jailbreak exploit in Fable 5—underscores the escalating cat-and-mouse game between safety researchers and malicious actors, and the government’s willingness to treat even a single red-team finding as grounds for an industry-wide intervention. The fact that Mythos 5 shares the same underlying model as Fable 5 made the suspension sweeping, but the subsequent differentiation—keeping the general-purpose version banned while releasing the specialized one to cyber defenders—hints at an emerging framework where an AI’s intended use case determines its regulatory fate.

For the AI industry, this episode carries profound implications. It sends a clear signal that the most capable models may only be deployed to entities deemed “trusted” by the government, creating a bifurcation between vetted incumbents and the wider market. The absence of published criteria for inclusion mystifies the process and could concentrate power among a handful of defense-oriented companies, while startups and smaller enterprises risk being shut out of the latest capabilities. The parallel move by OpenAI to similarly limit GPT-5.6 suggests this is not an isolated incident but a de facto policy for frontier labs. Investors will need to price in regulatory risk and government-relations competence as essential valuation factors, possibly shifting capital toward firms with existing Beltway ties.