Anthropic restored global access to its Claude Fable 5 model after the U.S. Department of Commerce withdrew export controls imposed on June 12. The company implemented a specific safety filter to block prompts that identify software vulnerabilities and generate exploit code, satisfying requirements set by the Center for AI Standards and Innovation (CAISI).
The 18-day blackout ended the standoff between the AI lab and federal regulators. Because the June 12 directive barred any foreign national—including Anthropic’s own non-citizen employees—from accessing Fable 5 or its parent model, Mythos 5, the company lacked a reliable method to verify user nationality. Anthropic pulled both models worldwide to ensure compliance.
How the New Safety Classifier Blocks Exploits
The restoration hinges on a new classifier designed to detect a specific prompting technique flagged by Amazon researchers. These researchers discovered a method to force Fable 5 to identify software vulnerabilities and, in one instance, write functional code to exploit them. Anthropic’s new filter blocks this specific technique in more than 99% of cases.
Crucially, the filter acts as a traffic controller rather than a lobotomy. The model’s underlying capability to identify vulnerabilities remains intact; the classifier simply detects the request and reroutes the user to the older Opus 4.8 model. This detection-based approach is a reactive measure. Anthropic conceded that no model is fully robust against jailbreaks and expects more bypasses to surface.
This “rerouting” strategy creates a performance trade-off. Anthropic noted that the classifier occasionally catches benign coding and debugging requests, treating them as false positives and diverting them to the less capable Opus 4.8.
Why the “Mythos-Class” Cyber Threat Was Oversold
The export controls were driven by fears that “frontier” models possessed unprecedented cyber-offensive capabilities. However, a joint review conducted by Anthropic, Amazon, and the U.S. government suggested these capabilities are widespread across the industry, not unique to Fable 5.
- Cross-Model Parity: The review found that OpenAI’s GPT-5.5 and China’s Kimi K2.7 could identify the same vulnerabilities as Fable 5.
- Exploit Replication: Every model tested—including the lightweight Haiku 4.5 and Sonnet 4.6—could reproduce the single exploit demonstration found by Amazon.
- The Verdict: These findings backed the argument that the specific cyber capabilities of the Mythos-class models were exaggerated.
While Fable 5 is back for the general public, Mythos 5 remains restricted. It returned to a limited set of U.S. organizations on June 26 and continues to be available only to Project Glasswing partners due to its fewer guardrails.
The Battle for Benchmark Dominance
The outage created a vacuum in the AI leaderboard. During the 18-day absence of Fable 5, Z.ai’s GLM-5.2, a model from a Chinese lab, claimed several top positions by default. Most notably, GLM-5.2 held the top accessible score on the AA-Briefcase multi-week task test.
By restoring Fable 5, Anthropic reclaims its position in the high-end LLM market. The model is now available across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. Deployment on AWS, Google Cloud, and Microsoft Foundry is scheduled to follow.
Usage Credits and Access:
- Pro, Max, Team, and Enterprise Plans: Fable 5 usage counts toward only 50% of weekly limits through July 7.
- Post-July 7: The model will transition to a standard usage credit system.
- Security Feedback: Anthropic has launched a HackerOne program to incentivize researchers to find and report new jailbreaks.
What This Means for Future Model Releases
This episode establishes a new precedent for how the U.S. government interacts with frontier AI labs. Anthropic has committed to giving designated government partners earlier access to test future models before they are released to the public.
This “pre-clearance” model aims to prevent the kind of sudden, global shutdowns seen in June. By integrating CAISI reviews into the development cycle, Anthropic hopes to avoid the binary choice between a total blackout and a potential national security risk. However, the reliance on classifiers to block “known techniques” suggests a perpetual game of cat-and-mouse between developers and the CVE-tracking cybersecurity community.
For developers and enterprise users, the return of Fable 5 restores a critical tool for complex reasoning and long-context window tasks, provided their prompts don’t trigger the vulnerability filter.
