Anthropic has moved to restore public access to its most capable AI models, Claude Fable 5 and Mythos 5, after the US government ordered their temporary shutdown over cybersecurity concerns. The models were taken offline in mid-June following reports that researchers found ways to bypass safeguards, prompting Anthropic to suspend public redeployment while controls were tightened.
According to posts from Anthropic on Wednesday, the government lifted the restrictions and the company is redeploying Fable 5 with an updated safety approach. Anthropic said it is using “a new set of classifiers” designed to better target and block more cybersecurity tasks, reflecting a shift from broad access to narrower, more selectively controlled operation.
Key takeaways
- Anthropic says US export-related restrictions on Claude Fable 5 and Mythos 5 were lifted on Wednesday, enabling renewed public access.
- The models were pulled after reports of bypassing Fable 5’s safeguards that could potentially help identify vulnerabilities for malicious use.
- Anthropic plans to redeploy with revised cybersecurity safety classifiers aimed at blocking additional harmful task types.
- US officials emphasized “safe” deployment while reviewing model alignment with government priorities.
- Anthropic is also working on a broader framework to assess jailbreak severity through its Project Glasswing partnership network.
Why Anthropic’s access was paused
Public access to Anthropic’s two latest flagship models—Claude Fable 5 and Mythos 5—was restricted starting June 12. The immediate trigger, Anthropic indicated, was an export-control directive tied to government review of a report describing how researchers could bypass safeguards on Fable 5.
In that report, the bypass method allegedly allowed the model to surface multiple software vulnerabilities. Anthropic responded by pulling access to the models rather than leaving the public-facing system in a state the government considered too risky. The broader implication was that frontier model capabilities, when paired with inadequate resistance to misuse, could translate quickly into real-world security threats.
What changed in the redeployed model
Anthropic said the suspension is ending after “productive conversations with the US government.” In its statement, the company framed the update as an operational safety upgrade rather than a complete redesign: the redeployed model will use “a new set of classifiers” intended to detect and block more cybersecurity tasks.
This matters for users and developers because classifier-based controls directly affect what kinds of requests the model will refuse or redirect. In practical terms, those controls can determine whether legitimate security analysis workflows remain usable while high-risk directions are filtered more aggressively.
Anthropic also addressed a key concern raised during the shutdown: the company argued that the issue wasn’t uniquely tied to Fable 5. In a blog post titled “Redeploying Fable 5,” Anthropic said weaker models could potentially identify similar vulnerabilities and produce the same exploit pathways. That framing suggests the company views the problem as a category risk across model capability levels, not a single-model anomaly.
US government concerns and the push for rapid deployment
US Secretary of Commerce Howard Lutnick said on X that, over the past two weeks, the government worked closely with Anthropic to “analyze and approve Fable 5,” aiming for alignment across the US government and to strengthen American AI leadership. Separately, White House Chief of Staff Susie Wiles said the priority remained to “get the best tech deployed as quickly and safely as possible.”
While those remarks do not change the underlying technical issues, they highlight the policy balance driving the intervention: the US government wants frontier models available, but only under conditions it believes reduce the chance of misuse—particularly misuse involving cybersecurity.
The White House concern, according to the article’s context, centered on the possibility of jailbreaking: if powerful models can be coerced into producing harmful instructions, they could become a national security risk. That is the logic behind export-control-style restrictions for models deemed high-impact and harder to contain once broadly available.
Anthropic’s Project Glasswing and a new approach to jailbreak severity
The model shutdown also cast attention on how AI safety should be tested and governed beyond one-off fixes. In the wake of the restrictions, Anthropic said it has begun drafting a consensus framework with partners including Amazon, Microsoft, Google and others for assessing the severity of AI jailbreaks.
The effort is connected to Project Glasswing, a collaboration announced in April focused on safeguarding against AI cybersecurity threats. A key goal appears to be establishing a shared way to evaluate jailbreak attempts: not just whether a safety system can be bypassed, but how dangerous the bypass outcome is and how consistently it can be reproduced.
Anthropic also said it is scaling up collaboration with the US government on model testing and safeguards. The company described plans that include pre-release access to models and safeguards for evaluation, information sharing about jailbreaks and misuse, and dedicated resources for joint research.
In parallel, an AI researcher previously claimed to have jailbroken Fable 5 within 48 hours of its June launch, before the government restrictions were applied. The researcher shared screenshots showing how the guardrails were allegedly bypassed. While that claim is reported as part of the surrounding context, it underscores why governments and model providers view guardrail testing as an ongoing race rather than a one-time checkpoint.
For the broader AI market, these developments raise a familiar tension: the same capabilities that make frontier models useful also raise the cost of safety failures, especially when cybersecurity misuse is possible. The redeployment suggests Anthropic believes the controls can be improved quickly enough to keep pace, but the existence of an industry-wide framework proposal indicates the challenge is bigger than one company’s deployment decisions.
Going forward, readers should watch for how Anthropic’s classifier changes alter real user behavior—what kinds of cybersecurity requests remain available, which patterns are newly blocked, and whether the proposed jailbreak-severity framework results in more transparent testing standards across major model providers.
https://www.cryptobreaking.com/anthropic-plans-to-resume-fable/?utm_source=blogger%20&utm_medium=social_auto&utm_campaign=Anthropic%20Plans%20to%20Resume%20Fable%205%20as%20US%20Eases%20Export%20Controls%20
Comments
Post a Comment