By 
International AI player, Anthropic PBC is initiating a wide release of its highly anticipated AI model under a dual-track strategy designed to balance commercial capabilities with critical safety guardrails. The company announced it will roll out a restricted version called Fable 5 on Tuesday, months after warning that its underlying powerful model could autonomously identify and exploit software vulnerabilities.
Fable 5 features strict guardrails preventing it from addressing cybersecurity and biology queries. When users input prompts in these sensitive domains, Anthropic’s Claude chatbot will automatically reroute the queries to an alternative model, Opus 4.8.
Concurrently, Anthropic is offering the unrestricted, cyber-capable version under a different moniker: Mythos 5. This version remains exclusively available to select entities via “Project Glasswing.”
Citing risks that the model can exploit vulnerabilities across every major operating system and web browser, Anthropic has heavily restricted access. However, the company recently added 150 organizations to Project Glasswing, expanding the trusted group to roughly 200 firms.
“We wanted to make sure for non-cyber use cases, we really prioritized safely releasing Fable as soon as possible,” said Dianne Penn, Anthropic’s head of project management for research and labs, noting that the company continues to refine safety protocols for broader cyber applications.
Despite the restrictions, Fable 5 boasts significant upgrades in coding, financial analysis, and long-horizon problem-solving. In a pilot test, payment processor Stripe utilized the software to complete a complex engineering task in a single day, a feat Anthropic estimates would normally require a human team two months to execute manually. Additionally, a biological hypothesis generated by the model regarding an E. coli protein mechanism was recently validated by an independent research laboratory.
To ensure the integrity of Fable 5’s guardrails, Anthropic subjected the model to an external red-teaming bug bounty program. Over more than 1,000 hours of testing, researchers failed to discover any universal jailbreaks. Penn emphasized that Anthropic will continue expanding access to the cyber-capable Mythos 5 through Project Glasswing to trusted partners moving forward.
Comments