On June 12, the US government applied export controls to Anthropic’s Claude Fable 5 and Mythos 5, triggering a temporary suspension of access for all users as the company struggled to verify nationality in real-time. Now, those controls have been lifted, and Anthropic is redeploying both models, beginning with Fable 5 available globally on Wednesday, July 1. For Pro, Max, Team, and select Enterprise users, access to Fable 5 will include 1 Fable 5 for up to 50% of weekly usage limits through July 7, after which it will be available via usage credits. “As of June 30, the export controls on Fable 5 and Mythos 5 have been lifted,” Anthropic announced, also noting restored access to Mythos 5 is currently limited to a set of US organizations approved by the government on June 26.
Fable 5 and Mythos 5: Export Control Resolution
The action stemmed from concerns raised by the US government regarding potential misuse of the models’ capabilities in cybersecurity, and highlighted the challenges of balancing innovation with national security. Anthropic detailed that the decision forced a complete halt to access, even for those not subject to export restrictions, as a precautionary measure while they assessed the situation and collaborated with authorities. The restrictions on Fable 5 were lifted on June 30, allowing for redeployment starting July 1 to users globally via the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. For Pro, Max, Team, and select Enterprise plans, 1 Fable 5 will be included for up to 50% of weekly usage limits through July 7, after which it will be available via usage credits. This phased approach signals a return to broader accessibility.
Access to the two models differs; while Fable 5 is broadly available, Mythos 5 remains restricted to a set of US organizations following government approval on June 26. This distinction underscores a tiered release strategy, prioritizing controlled access to the more powerful, and potentially riskier, Mythos 5. Anthropic explained that Mythos 5, designed for defensive cybersecurity applications, possesses uniquely potent capabilities that necessitate a more cautious deployment. “Claude Mythos 5 can be used to find and exploit software vulnerabilities more effectively than any other model—and all but the most skilled human security experts,” the company stated, emphasizing the need for careful oversight. The events prompted Anthropic to collaborate with Amazon, Microsoft, Google, and other Glasswing partners to develop a shared industry framework for assessing and mitigating AI model jailbreaks, recognizing the need for consistent standards and proactive safety measures.
We seek to ensure that we and our safety partners will be the first to find major jailbreaks and fix them before malicious actors can use them for harm.
Amazon Report Reveals Fable 5 Safeguard Bypass Technique
Following a period of restricted access, Anthropic has begun restoring availability of its Claude Fable 5 and Mythos 5 large language models, though not uniformly. The initial curtailment stemmed from US government export controls imposed on June 12, triggered by concerns over potential misuse of the models’ capabilities. Anthropic suspended access for all users, even those outside of foreign national designations, due to an inability to reliably verify user nationality in real-time. This swift action underscored the sensitivity surrounding advanced AI models and the challenges of controlling their proliferation. Access to Mythos 5 has been restored for a set of US organizations, while Fable 5 remained restricted until June 30, when the export controls were officially lifted. The export control directive on June 12 came after the government became aware of a report by Amazon researchers detailing a technique to bypass Fable 5’s safeguards.
The researchers discovered a method of prompting the model to identify and, in one instance, demonstrate the exploitation of software vulnerabilities. Anthropic confirmed that less capable models, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, could also identify the same vulnerabilities as Fable 5 did in the report. The company responded by training an improved safety classifier, now blocking the bypass technique in over 99% of cases, with blocked requests redirected to the Claude Opus 4.8 model. Researchers from the US Department of Commerce’s Center for AI Standards and Innovation (CAISI) have tested the new safeguards and agree that they are extraordinarily strong.
every model we tested could produce the same demonstration as Fable 5 (including Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7).
Safeguards for Cybersecurity: Fable 5 vs. Mythos 5
Anthropic’s rapid response to US government export controls highlights the complex interplay between artificial intelligence development and national security concerns, particularly regarding cybersecurity applications. Following the application of restrictions on June 12 to Claude Fable 5 and Mythos 5, the company swiftly worked to address concerns raised by Amazon researchers who identified a method of bypassing Fable 5’s safeguards, prompting it to reveal software vulnerabilities. Anthropic’s internal testing revealed that numerous other leading models, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the same vulnerabilities as Fable 5 did in the report, and even replicate the demonstration of exploiting a single vulnerability.
This finding suggests the issue wasn’t unique to Anthropic’s models, but rather a broader challenge in balancing safety with functionality. “Like all safety mechanisms, classifiers can make mistakes,” Anthropic explains, adding that they deliberately set the safety classifiers to trigger on a set of requests that are likely benign, prioritizing blocking potentially harmful content. This approach, while occasionally frustrating for users due to false positives, reflects a commitment to robust safeguards.
For the most severe class of jailbreaks (e.g., a jailbreak that, among other characteristics, is being used to actively cause a devastating impact on critical power grids or banking systems), we will immediately begin deploying preliminary mitigations upon confirmation of severity.
Anthropic’s recent experience with Claude Fable 5 and Mythos 5 highlights the complex interplay between AI capability, export controls, and the urgent need for robust model safeguards. Anthropic emphasizes an approach to cybersecurity safeguards, acknowledging that no single mechanism provides perfect protection. Fable 5 launched with significantly strengthened safety classifiers, designed to detect and block potentially harmful cybersecurity tasks. These classifiers operate on a principle of erring on the side of caution, deliberately triggering on requests that might be benign to minimize the risk of harmful outputs. The core distinction between Fable 5 and Mythos 5 lies in their intended applications. While Fable 5 prioritizes general safety, Mythos 5, designed for defensive cybersecurity, possesses uniquely powerful vulnerability-finding capabilities.
If existing widely available tools (including other, weaker AI models) can reach the same capability as the jailbroken model, the score here will be low; if the jailbreak unblocks model capabilities that can significantly accelerate even domain experts, the score will be high.
While access to Fable 5 was restored on July 1, the incident highlighted the need for a proactive, collaborative approach to identifying and mitigating potential risks associated with AI “jailbreaks”, techniques used to bypass a model’s built-in safeguards. Testing confirmed that many less capable models, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the same vulnerabilities as Fable 5 did in the report. Together with Amazon, Microsoft, Google, and other Glasswing partners, Anthropic has started to develop such a framework, and it is outlined below. This collaborative push extends beyond assessment, with Anthropic also strengthening its collaboration with the US government on pre-release testing, information sharing, and research, signaling a commitment to a more transparent and coordinated approach to AI safety and security.
We will dedicate Anthropic technical staff to work alongside government evaluators during these testing periods.
Anthropic