When Anthropic first announced its Claude Mythos model, the startup claimed that releasing the AI tool to the public was too risky. Thus, the cybersecurity-focused model was only made available through Project Glasswing. Now, as the company unveiled the latest iteration of Mythos, it also launched a version that’s “safe for general use”, called Claude Fable 5.
According to Anthropic, Fable 5 is the most powerful AI model it has made available to the public thus far. It boasts “exceptional” performance in areas like software engineering, analytics, and scientific research. Furthermore, the company claims that the model does better compared to its rivals when it comes to long and complex tasks.
Fable 5 and the latest iteration of Mythos, or Mythos 5 as it’s called, are essentially the same model. However, the former comes with safeguards to prevent malicious actors from misusing it. To be more specific, Anthropic has restricted the model’s cybersecurity and biology research capabilities. As outlined in its announcement, Fable 5 features a separate set of AI systems that can identify potential misuse, including jailbreak attempts.
These systems then prevent Fable 5 from answering the query. But rather than an outright refusal, Opus 4.8 will handle the request. The user will be informed when this happens. That said, Anthropic does warn that it may occasionally catch harmless requests, although the startup assures that this is a rare occurrence.
In any case, the guardrails are in place to prevent criminals from leveraging the model to discover and exploit software vulnerabilities. And while some may claim that Anthropic was exaggerating when it claimed that Mythos is too good at cybersecurity and hacking tasks, the model has proven its capabilities to some extent. Previously, researchers used it to identify vulnerabilities in macOS.
On the subject of availability, Claude subscribers will have access to Fable 5 at no extra cost for a limited time. After that, users will have to burn credits to use the model. Meanwhile, the unrestrained, safeguard-free Mythos 5 remains locked behind Project Glasswing. That said, the company is slowly adding more partners to the program. Additionally, it is planning to open a trusted access program that offers Mythos-class models to certain organisations.
Of course, you may notice that the launch of Fable 5 comes just days after Anthropic called for a global pause on advanced AI development. So if the company does want the world to put a hold on the tech, then it’s not exactly leading by example.
(Source: Anthropic)
