Anthropic blocks customer access to Fable 5 and Mythos 5 over national security concerns

At a glance:

• Anthropic disabled access to Fable 5 and Mythos 5 models on June 12 following a US government directive.
• The action stems from concerns about a potential jailbreak method compromising cybersecurity capabilities.
• The company criticized the government's approach as non-transparent and misaligned with AI oversight principles.

Government directive and immediate response

Anthropic announced on June 12 that it had fully disabled customer access to its Fable 5 and Mythos 5 AI models to comply with a US government directive citing national security risks. The move affects all users, including foreign nationals both within and outside the US, as well as Anthropic employees. The company emphasized that its other models and the Claude chatbot remain unaffected by the restriction. This action follows the June 9 launch of Fable, a public-facing AI model designed to extend Mythos' advanced capabilities, which Anthropic claims surpass previous iterations in performance and versatility.

The US government did not publicly disclose the specific nature of the security concerns prompting the order. However, Anthropic suggested the directive was triggered by reports of a jailbreak method targeting Fable 5. The company noted that while safeguards were implemented to mitigate misuse risks—particularly in cybersecurity applications—the government's intervention bypassed its preferred framework for addressing unsafe AI deployments. Anthropic has historically advocated for transparent, technically grounded regulatory processes for AI oversight.

Fable 5's capabilities and testing benchmarks

Fable 5 was introduced as a successor to Mythos, a cybersecurity-focused model restricted to Anthropic's Project Glasswing partners. The public version of Fable demonstrated superior performance compared to earlier models, including successfully completing tasks that previous versions could not. For example, during internal testing, Fable 5 reportedly achieved higher scores in complex problem-solving benchmarks than its predecessors. Notably, it outperformed the original Pokémon Red game in a test scenario where Claude, the company's general-purpose model, had previously failed to surpass the game's challenges.

Despite these advancements, Anthropic acknowledged that no AI model can guarantee absolute immunity against jailbreak attempts. The company stated that its safeguards were designed to make such exploits either highly specific (non-universal) or resource-intensive to develop. It also emphasized ongoing monitoring systems intended to detect and neutralize successful jailbreaks rapidly. However, user feedback indicated that some safeguards were overly restrictive, limiting legitimate use cases.

Government concerns and jailbreak risks

The US government's directive was based on verbal evidence of a potential narrow jailbreak affecting Fable 5, according to Anthropic. The company did not identify the entity that provided the evidence but noted that the alleged vulnerability was not universal, meaning it would not apply to all users or scenarios. Anthropic expressed disagreement with the government's decision to restrict access, arguing that potential vulnerabilities should not automatically justify blocking commercial models. The company pledged to release additional technical details within 24 hours of the announcement.

Anthropic's statement highlighted its belief in a statutory framework for AI governance that prioritizes transparency, fairness, and technical accuracy. It criticized the current directive as inconsistent with these principles, suggesting that the government's approach lacked the procedural rigor it has publicly advocated for. The company's stance reflects ongoing tensions between AI developers and regulators over balancing innovation with security risks.

Regulatory implications and future outlook

This incident underscores growing scrutiny of AI models with dual-use potential, particularly in cybersecurity and national security contexts. Anthropic's compliance with the directive signals the increasing influence of government oversight on AI development and deployment. However, the company's public criticism of the process raises questions about how such interventions will be managed in the future.

The restriction on Fable 5 and Mythos 5 may impact Anthropic's commercial strategy, especially if similar directives are issued for other models. The company's emphasis on safeguards and monitoring could become a template for addressing regulatory concerns without outright model restrictions. Meanwhile, the broader AI community is likely to watch how this case influences policies around model access and security protocols.

Industry reactions and broader context

The move comes amid heightened global discussions about AI regulation, with governments seeking to balance innovation and security. Anthropic's situation mirrors challenges faced by other AI developers navigating compliance with evolving national security mandates. The company's disagreement with the government's approach may resonate with peers advocating for clearer, collaborative frameworks for addressing risks.

Project Glasswing partners, who had access to Mythos 5 prior to the restriction, may face uncertainties about future access to advanced models. The incident also highlights the complexity of securing AI systems against adversarial attacks, a challenge that extends beyond Anthropic to the entire industry. As AI capabilities advance, such dilemmas are expected to intensify, shaping both technological and regulatory trajectories.