Amazon Found Fable 5 Bypass; Mythos 5 Still Restricted to 30% of U.S. Orgs

For cybersecurity professionals, the temporary freeze and partial thaw of Anthropic's Claude models is more than a regulatory drama: it is a real-world stress test of AI's offensive cyber capabilities. The discovery by Amazon researchers that Fable 5's safeguards could be bypassed to find and exploit software vulnerabilities transforms frontier models from abstract risks into concrete tools that adversaries could wield against critical infrastructure. With Mythos 5—expressly noted for its vulnerability-finding prowess—still restricted to a government-vetted pool, defenders must now operate under the assumption that AI-accelerated attacks are not a future scenario but a present danger.

The Trump administration's lifting of restrictions on Anthropic's Claude models marks a significant pivot in the uneasy dance between AI innovation and national security. On June 12, the Commerce Department had blocked foreign nationals from accessing the company's latest models—Claude Fable 5 and the more potent Mythos 5—after cybersecurity researchers at Amazon, Anthropic's primary cloud provider, found a method to bypass Fable 5's safeguards. The exploit allowed the AI to discover and potentially exploit software vulnerabilities, a capability that authorities feared could be weaponized by malicious actors against critical infrastructure.

“

The Trump administration's lifting of restrictions on Anthropic's Claude models marks a significant pivot in the uneasy dance between AI innovation and national security.

Now, weeks later, the administration has relented, albeit conditionally. Fable 5 is once again widely available, but Mythos 5 remains tightly restricted: only a hand-picked group of U.S.-based organizations approved by the federal government gain access. This graduated approach reflects a broader recalibration in how the U.S. evaluates frontier AI. The timing is notable: just last month, President Trump signed an executive order establishing a framework for the federal government to review the national security risks of the most advanced AI systems for up to 30 days before their public release. While participation is technically voluntary, the dual-track reinstatement of Anthropic's models—and the simultaneous announcement that OpenAI will restrict its own GPT-5.6 Sol to a similar government-vetted pool—suggests the framework is already exerting coercive pressure.

The immediate catalyst, the Amazon cybersecurity report, underscores a critical vulnerability residing not in the models themselves but in the safeguards meant to contain them. Bypassing Fable 5's guardrails essentially made it a zero-day discovery engine. The potential for abuse is stark: a state-sponsored group could use a model like Mythos 5 to scan power-grid software, communication protocols, or financial systems, identifying entry points far faster than human teams. This is precisely why the administration's initial reaction was a blanket ban on foreign access. The targeting of foreign nationals—a blunt instrument—highlighted the administration's preoccupation with preventing model weights or outputs from leaking to rival states, a concern amplified by Anthropic's own earlier warnings that Mythos was adept at finding software flaws.

The partial lift now signals that the 30-day review window, still under development, has produced some level of reassurance. For Fable 5, the government appears satisfied that the safeguard bypass has been patched or that the model's capabilities do not meet a catastrophic threshold. Mythos 5, however, remains in a liminal state, accessible only to U.S. entities vetted by the government. This hints at a tiered classification system emerging: models with high but not extreme offensive potential can be released widely; those with the power to systematically undermine cybersecurity at scale require ongoing surveillance.

For the cybersecurity industry, this episode injects new urgency into the conversation around AI red-teaming and automated vulnerability discovery. It underscores that external safety evals—especially from a model's own cloud provider—are becoming a prerequisite for government approval. Amazon's role here is pivotal; it acted as a de facto auditor, perhaps unintentionally setting a precedent that cloud hyperscalers could be compelled to police the AI workloads running on their infrastructure. This could strain commercial relationships and raise antitrust questions, even as it creates a de facto security layer.