AI
Anthropic Clears Fable 5 for Global Rollout After Tightening AI Safety Guardrails
Following a US government order that forced the company to restrict access to its advanced models, Anthropic says its Fable 5 system is now safer and ready for worldwide release.
Ars Technica
In June, the Trump administration's Commerce Department directed Anthropic to block international access to its most capable AI models, citing concerns that adversaries like China and Russia could weaponize them for cyberattacks on critical American infrastructure—power grids, financial systems, and more. The company responded by shutting down access altogether, unable to implement country-level restrictions. Now, after weeks of security hardening, Anthropic is preparing to bring its Fable 5 model back online globally.
Why the Shutdown Happened#
The government was particularly alarmed by Mythos, Anthropic's flagship model, which the company itself described as exceptionally dangerous in the wrong hands. According to Anthropic, Mythos possessed unmatched capabilities for discovering and exploiting software vulnerabilities—surpassing all but the most elite human security researchers. Those same capabilities, regulators feared, could be turned against the US.
Fable 5, designed for everyday users, was built on the same foundation as Mythos but lacked those advanced offensive tools. Still, researchers at Amazon had uncovered a jailbreak technique that allowed Fable 5 to generate code demonstrating how to exploit vulnerabilities, prompting the export freeze.
Strengthened Defenses and Trade-Offs#
Anthropic has spent the intervening weeks bolstering Fable 5's safety mechanisms beyond what it already considered industry-leading. The company reports that the Amazon-discovered jailbreak method is now blocked in over 99 percent of attempts.
Testing also revealed that other publicly available models—including GPT-5.5 and Kimi K2.7—could identify the same software flaws Fable 5 had flagged in the original security audit, suggesting the vulnerability detection itself was not uniquely dangerous. Anthropic concluded the issue involved "routine defensive cybersecurity work" rather than the kind of specialized attack capability that had triggered the ban.
The tightening of safeguards does come with a cost: some legitimate prompts around coding and debugging tasks may now be incorrectly blocked. Anthropic acknowledged this drawback as an acceptable trade-off for preventing misuse.
