Anthropic Restores Fable 5 With Tightened Safeguards

Anthropic wrote in a June 30 blog post that it is redeploying Claude Fable 5, available globally July 1 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork (Anthropic blog post). Reuters and CNBC report the U.S. Commerce Department lifted export controls it had imposed on June 12, which had required Anthropic to cut off Fable 5 and Mythos 5 access for foreign nationals, including the company's own non-citizen staff. The controls followed Amazon researchers reporting to U.S. officials that they had found a way to prompt Fable 5 into identifying software vulnerabilities and, in one case, writing exploit code, according to Tom's Hardware and 9to5Google.

Axios and CNBC report Anthropic's fix was a new classifier layer tuned specifically to the flagged jailbreak technique, which the company says blocks it about 99% of the time in testing, while acknowledging the tighter filtering can also flag benign cybersecurity or coding queries. Reuters reports Commerce's Center for AI Standards and Innovation (CAISI) reviewed the safeguards before controls were lifted. 9to5Google and Anthropic's blog post note that during testing, similar outputs were reproduced on other models, including Opus 4.8 and GPT-5.5, suggesting the underlying exploit pattern is not unique to Fable 5's architecture. In exchange for restored access, Anthropic has agreed to proactively detect and address security risks, help develop standards for future models, and report malicious activity to the government, per reporting on the resolution.

• •Deployments of frontier models can be interrupted by government action informed by third-party red-team findings, not just vendor-initiated safety reviews (Reuters; 9to5Google).
• •Narrow, technique-specific classifiers can resolve a flagged exploit faster than a full retrain, but they trade some false-positive risk on legitimate security and coding workflows (Axios; Anthropic blog post).
• •Cross-model reproduction of an exploit pattern (Opus 4.8, GPT-5.5) signals teams should treat this class of vulnerability as an industry-wide risk, not vendor-specific.

Whether Commerce's CAISI publishes standardized testing criteria for future frontier releases, how quickly access returns on AWS, Google Cloud, and Microsoft Foundry marketplaces (CNBC), and whether Anthropic's commitment to report malicious activity to the government becomes a template other frontier labs adopt under future export-control reviews.