Table of Contents
Anthropic says it has been ordered by the US government to immediately suspend access to its flagship AI models, Fable 5 and Mythos 5, for all customers worldwide. According to the company, the directive was issued under national security authorities and applies to all foreign nationals, including Anthropic employees who are not US citizens.
The company strongly disagrees with the decision, arguing that the reported security issue involves only a narrow jailbreak technique that does not provide capabilities beyond what many other publicly available AI models already offer. While Anthropic is complying with the order, it warns that applying this standard across the industry could significantly slow or even halt future frontier AI deployments.
The move raises major questions about AI regulation, model safety, government oversight, and the future relationship between AI companies and national security agencies.
Anthropic Says Government Issued Sudden Suspension Order
According to Anthropic, the company received the directive at 5:21 PM ET and was instructed to disable access to both Fable 5 and Mythos 5 immediately.
The government reportedly cited national security authorities but did not provide detailed information about the underlying concern.
Anthropic claims its understanding is that officials became aware of a technique capable of partially bypassing, or “jailbreaking,” Fable 5’s safety protections.
A jailbreak refers to a method that attempts to circumvent an AI model’s built-in safeguards, allowing users to obtain responses that would normally be restricted.
However, Anthropic argues that the vulnerabilities demonstrated through this technique were already known, relatively minor, and discoverable by other AI systems as well.
As a result, the company says it was surprised by the severity of the government’s response.
Why the Government May Have Acted
While officials have not publicly explained the directive, several possible factors may have influenced the decision.
1. Growing National Security Concerns Around AI
Governments worldwide are increasingly worried that advanced AI systems could be used for:
- Cybersecurity attacks
- Vulnerability discovery
- Malware development
- Intelligence gathering
- Critical infrastructure targeting
Even if a model’s safeguards are strong, regulators may view any successful jailbreak as evidence that additional restrictions are needed.
2. Fear of Capability Escalation
Frontier AI models are becoming increasingly capable of analyzing large codebases, identifying vulnerabilities, and assisting with software engineering tasks.
Authorities may be concerned that even narrow jailbreak techniques could eventually evolve into broader methods that expose more advanced capabilities.
3. Precedent Setting
Governments are still determining how to regulate highly capable AI systems.
By taking a strong position now, regulators may be attempting to establish a precedent that advanced AI providers must meet exceptionally high safety standards before deployment.
Anthropic’s Defense: Fable 5 Was Extensively Tested
Anthropic strongly pushed back against the decision in its statement.
The company says Fable 5 underwent extensive red-team testing before launch, involving:
- Internal security teams
- Third-party security organizations
- UK AI Safety Institute participation
- Collaboration with US government entities
According to Anthropic, thousands of hours of testing were conducted to evaluate the model’s resilience against misuse.
The company claims these tests demonstrated that Fable’s safeguards were significantly stronger than those found in previous generations of AI models.
Most notably, Anthropic says researchers were unable to identify a universal jailbreak capable of broadly bypassing safety protections.
Instead, the company says only narrow and limited jailbreaks have been observed.
Anthropic’s “Defense in Depth” Strategy
One of the most interesting aspects of the company’s response is its emphasis on a defense-in-depth security model.
Rather than claiming jailbreaks are impossible, Anthropic openly acknowledges that no AI provider has solved the jailbreak problem completely.
The company argues that perfect resistance is currently unrealistic for any large language model.
Instead, its strategy involves multiple layers of protection:
- Strong safety training
- Behavioral restrictions
- Continuous monitoring
- Threat detection systems
- Logging and investigation capabilities
- Rapid mitigation of discovered exploits
Anthropic says this layered approach is intended to make jailbreaks difficult, expensive, and limited in scope.
In practical terms, the company argues that reducing risk is more realistic than attempting to eliminate it entirely.
The Controversial 30-Day Data Retention Policy
Anthropic also highlighted a policy that generated significant debate when Fable 5 launched.
The company requires customer data retention for 30 days when using Fable 5.
At the time, some users criticized the policy as intrusive.
Anthropic now argues that the retention period exists partly to help detect and investigate jailbreak attempts.
According to the company, maintaining access logs allows engineers to identify emerging attack techniques and deploy fixes more quickly.
The statement suggests Anthropic views monitoring as a critical component of AI safety rather than relying solely on preventative safeguards.
Why Anthropic Believes the Suspension Is Excessive
Anthropic’s central argument is straightforward.
The company says the reported jailbreak:
- Was not universal
- Was limited in scope
- Produced only minor findings
- Did not demonstrate unique Mythos-specific capabilities
- Showed behavior available from other public AI models
The company specifically claims that the capability demonstrated in the reported findings is already available through competing systems, including OpenAI’s GPT-5.5.
From Anthropic’s perspective, recalling a commercial AI model used by millions because of a narrow jailbreak creates an unrealistic regulatory standard.
The company warns that if every frontier AI deployment faced suspension whenever a limited jailbreak emerged, innovation across the entire AI industry could slow dramatically.
What This Means for the AI Industry
Whether one agrees with Anthropic’s position or not, the situation highlights a larger issue facing AI developers.
As models become more capable, regulators face a difficult balancing act:
- Protecting national security
- Preventing misuse
- Supporting innovation
- Encouraging competition
- Creating predictable regulatory frameworks
Companies need clear rules for what constitutes an unacceptable risk.
Governments need confidence that advanced AI systems are not introducing new threats faster than safeguards can keep up.
The clash between Anthropic and regulators may become one of the earliest major tests of how these competing priorities are resolved.
What Happens Next?
Anthropic says it is fully complying with the directive while working to restore access as quickly as possible.
The company also indicated that it plans to share additional technical details regarding the alleged jailbreak and its findings.
Those disclosures could become important in determining whether the suspension represents a justified national security intervention or an example of regulatory overreach.
For now, access to Fable 5 and Mythos 5 remains suspended, while the broader AI industry watches closely.
The outcome could influence how governments evaluate and regulate future generations of advanced AI systems for years to come.