Redeploying Claude Fable 5

The export controls were originally applied on June 12, citing potential cybersecurity concerns after Amazon researchers discovered a method to bypass Fable 5's safeguards by prompting the model to identify software vulnerabilities. In one instance, the model produced code demonstrating exploitation of a specific vulnerability. Anthropic's own testing revealed that the technique was not unique to Fable 5. Claude Opus 4.8, GPT-5.5, and Kimi K2.7 could each identify the same vulnerabilities. Every model Anthropic tested, including Claude Haiku, Sonnet, and Opus variants, could produce the same demonstration. The technique did not expose any unique Mythos-level cyber capabilities and represented a borderline case for Fable 5's safeguards, not a novel attack vector specific to the model.

Anthropic's safeguard architecture uses a defense-in-depth model that combines multiple independent protection layers. Classifiers are smaller automated AI systems running alongside the main model to detect when it is asked to perform a potentially harmful cybersecurity task. These systems operate separately from the model's own judgment, giving Anthropic a distinct enforcement layer that can be retrained and updated faster than the base model itself. The improved classifier rolled out for the July 1 redeployment was trained specifically against the reported bypass technique, not as a broad update to the general safeguard suite. Teams using Fable 5 for security research, penetration testing tooling, or vulnerability management should test their specific prompts after redeployment, since the tighter classifier margin may flag some legitimate requests that previously went through without friction.

Anthropic is partnering with Amazon, Microsoft, Google, and other Project Glasswing participants to build a shared industry framework for assessing jailbreak severity. The framework scores exploits across four criteria: capability gain (how far beyond existing tools the jailbreak extends a user's abilities), breadth of capability gain (how many distinct offensive tasks the same technique unlocks), ease of weaponization (the human effort required to convert the jailbreak into a working attack), and discoverability (how easily an adversary could find and replicate the technique independently).

Those four scores map to three categories. Minor jailbreaks fall within acceptable safety margins. Narrow harmful jailbreaks unblock specific restricted behaviors. Universal jailbreaks unblock entire classes of harmful capability. Anthropic reports that no universal jailbreaks for Fable 5 had been discovered at the time of the redeployment announcement. The framework will be submitted to the US Department of Commerce's Center for AI Standards and Innovation and the Office of the National Cyber Director. The practical goal is a consistent public threshold for when developers, enterprises, and governments need to act, replacing the current situation where every organization applies its own severity judgment independently.

As part of the resolution, Anthropic has committed to four specific areas of ongoing government collaboration. First, designated government partners will receive pre-release access to new models and their safeguards, supported by dedicated Anthropic technical staff, for independent evaluation before public launch. Second, Anthropic will rapidly investigate and notify partners of any significant jailbreak discoveries, share new safeguards for independent testing, and provide advance threat intelligence before public disclosure. Third, the company will dedicate internal research teams to government priorities, allocate significant compute resources for joint security testing, and contribute safety expertise to interagency cybersecurity vulnerability databases. Fourth, Anthropic will work with its Glasswing partners toward shared, voluntary security and evaluation standards across the frontier model provider community.

For Pro, Max, Team, and select Enterprise customers, Fable 5 access through July 7 was included within existing weekly usage limits at up to 50% of those limits. After July 7, usage draws on credits in the same way as other premium models. Teams on AWS Bedrock, Google Cloud, or Microsoft Azure Foundry faced a separate re-enablement process through their respective cloud providers. Anthropic also launched a HackerOne bug bounty program specifically for Fable 5 cybersecurity jailbreak submissions, giving security researchers a formal, paid channel to report newly discovered bypass techniques directly to Anthropic.

The 18-day suspension is worth reviewing not just as a news event but as a template for how export controls are likely to work as frontier AI capabilities become increasingly dual-use. Anthropic was not accused of wrongdoing. The government responded to a specific reported vulnerability, and Anthropic resolved it with a measurable technical fix. That pattern, a specific incident followed by a specific classifier improvement rather than a broad restriction, is a more workable regulatory model than blanket bans. For enterprises building production workflows on Fable 5, the practical takeaway is to build for the possibility of short, unannounced model suspensions. Keep tested fallback paths to Claude Opus 4.8 or Claude Sonnet 5 in your production architecture so your pipelines can continue running without manual intervention if access is restricted again.