The US government has ordered Anthropic to disable Claude Fable 5 and Mythos 5 globally, citing jailbreak vulnerabilities. Anthropic is complying but contesting the decision publicly.
The company argues the flagged vulnerabilities are minor and present in competing models including OpenAI's GPT-5.5. Anthropic characterizes the shutdown as regulatory overreach that could establish a chilling precedent for all frontier model deployments.
This move carries particular irony. Anthropic spent months promoting the cybersecurity risks of its Mythos class models to justify their development and testing. The government's action transforms those same concerns into grounds for a complete worldwide shutdown.
Anthropic frames this as a proportionality problem. Rather than implement targeted mitigations or restricted access, the government opted for total disablement. The company warns this approach contradicts industry norms around responsible scaling and could deter other AI labs from transparent disclosure of their own security findings.
The regulatory action reflects growing tension between AI developers and government oversight bodies. Federal authorities have escalated scrutiny of frontier models, particularly those approaching or reaching advanced reasoning capabilities. Jailbreak vulnerabilities, where users bypass safety guidelines through prompt engineering or indirect techniques, represent a persistent regulatory concern.
What remains unclear is whether other labs face similar pressure on comparable models or whether this action targets Anthropic specifically. The differential treatment between Anthropic's models and GPT-5.5 suggests selective enforcement, though government officials have not publicly explained their reasoning.
Anthropic's public pushback differs from its typical approach. The company generally cooperates with regulators quietly. This visible objection indicates the shutdown creates meaningful business impact and sets precedent the company views as genuinely dangerous to the field's future.
The decision underscores a fundamental regulatory question: at what vulnerability threshold should governments halt commercial deployments.