Anthropic has launched Claude Mythos 5 for cybersecurity and Claude Fable 5 for general use, emphasizing safety and advanced capabilities.
Overview of New AI Models
Anthropic has unveiled Claude Mythos 5, a specialized AI model focused on cybersecurity and scientific research, alongside a more user-friendly version, Claude Fable 5.
Targeted Access for Mythos 5
Initially, Mythos 5 will be accessible only to select entities such as cybersecurity firms, critical infrastructure operators, government partners, and specific life sciences researchers. Anthropic has highlighted potential risks associated with this advanced AI model, stating that it has reached a point where it could pose significant dangers.
Project Glasswing and Initial Concerns
Previously, Anthropic had limited access to Mythos through its Project Glasswing initiative while evaluating the associated risks. Concerns arose earlier this year when a preview version of Mythos was reported to autonomously complete a complex corporate network intrusion exercise and identify numerous vulnerabilities in software applications.
“Presenting Claude Fable 5: a Mythos-class model that has been made safe for general use,” stated Anthropic.
Fable 5: A Safer Alternative
Claude Fable 5 offers enhanced safety features for general users, while still boasting impressive capabilities that surpass previous models. Anthropic has noted that Fable 5 excels in various AI capabilities, including software engineering and scientific research.
Performance Benchmarks
Both Mythos 5 and Fable 5 achieved high scores across multiple benchmarks, with Mythos 5 scoring 78% on ExploitBench for vulnerability detection and 88% on Terminal-Bench for coding tasks. These results indicate that the new models can handle more complex tasks than their predecessors.
Safeguards in Place
To ensure safety, Fable 5 redirects sensitive inquiries to Claude Opus 4.8, which is activated in less than 5% of sessions. Anthropic has emphasized the importance of safeguards to prevent misuse of Fable 5's capabilities, particularly in cybersecurity.
Broader Collaboration with the U.S. Government
In a shift in relations with the U.S. government, Anthropic aims to expand access to Mythos 5, involving new partners in its trusted access program focused on cybersecurity initiatives. Reports suggest that the company has even embedded engineers within the National Security Agency for this purpose.
Industry Response and Criticism
The launch of these models comes amid ongoing debates about the responsibilities of AI developers. While Anthropic believes that Mythos could help in identifying vulnerabilities, critics have accused the company of using fear-based marketing tactics.
