Anthropic announced the release of two new AI models, Claude Fable 5 and Claude Mythos 5, on Tuesday. The company claims these models have enhanced capabilities compared to the Mythos Preview model, which was initially released to a limited group of tech industry partners in April. The decision to limit access to Mythos 5 stems from concerns about potential misuse of its advanced capabilities, including the development of hacking tools that could compromise cybersecurity defenses.
Claude Fable 5, now publicly available, uses the same underlying model as Mythos 5 but includes 'guardrails' to block queries related to cybersecurity, biology, and chemistry. These requests will be redirected to an older model, Claude Opus 4.8, if the company suspects distillation attempts. Diane Penn, Anthropic’s head of product management, stated that the company has been refining its approach to managing Mythos’ capabilities since before its April release, with testing and user feedback shaping the current strategy. She emphasized that this approach was deemed the most viable and beneficial for users, even if it isn’t perfect for all cases.
Anthropic is also providing access to Claude Mythos 5 to Project Glasswing partners and select biology researchers, with plans to expand access in the future. The company highlighted that the release of these models has forced global tech firms and governments to strengthen their cybersecurity defenses. Anthropic noted that while it has conducted extensive testing, no universal jailbreaks were found in over 1,000 hours of red-teaming. The company remains cautious about the potential risks of releasing such powerful models to the public.
Source: wired
