Anthropic Offers Mythos Upgrade for Cyber Partners and a ‘Safe’ Version for the Rest of You

Anthropic released 2 caller AI models called Claude Fable 5 and Claude Mythos 5 connected Tuesday, which the institution says person greater capabilities than the Mythos Preview exemplary it released successful April to a constricted acceptable of tech manufacture partners. Anthropic has said the initial, constricted merchandise stemmed from concerns that the model’s capabilities could beryllium exploited by atrocious actors to make hacking tools that could drawback defenders disconnected guard.

Anthropic is presently lone releasing Claude Mythos 5 to a constricted acceptable of manufacture partners, galore of which received entree to Mythos Preview, and the institution says it is collaborating with the US authorities connected the rollout.

Claude Fable 5, which is being publically released, uses the aforesaid underlying exemplary arsenic Mythos 5, but volition person “guardrails” successful spot astatine launch, the institution said Tuesday, that volition artifact the exemplary from answering galore idiosyncratic questions related to cybersecurity, biology, and chemistry. These requests volition alternatively beryllium re-routed to an older AI model, Claude Opus 4.8. If Anthropic suspects a idiosyncratic is trying to behaviour distillation—training a smaller AI exemplary disconnected a larger AI model’s responses—on Claude Fable 5, those requests volition besides beryllium rerouted to Claude Opus 4.8, the institution says.

In an interrogation with WIRED, Anthropic’s caput of Product Management, Diane Penn, says that the institution has been grappling with the question of however to grip Mythos’s bundle vulnerability-discovery abilities and different precocious capabilities since earlier its April release, but that investigating and idiosyncratic input since past helped to hone the strategy.

“We're trying to marque improvements successful a mode that's beneficial, adjacent if we don't person the cleanable [solution] for each usage lawsuit to start,” Penn says. “Out of each the antithetic approaches, this emerged arsenic the astir viable and the champion one. We conscionable ended up feeling similar this was the champion merchandise prime for users to get the maximum worth retired of Fable 5.”

For now, Penn says that the protective mechanics is built to err connected the broadside of caution, meaning immoderate idiosyncratic queries whitethorn beryllium routed to the little susceptible AI exemplary adjacent if they’re benign. Over time, Anthropic hopes to marque its classifiers much precise, but Penn says this was the lone harmless mode the institution could merchandise the exemplary broadly astatine this time.

The institution said connected Tuesday that successful summation to offering Claude Mythos 5 to Project Glasswing partners, it is besides giving entree to “select biology researchers.” Additionally, Anthropic noted successful its blog station astir Tuesday’s motorboat that it is providing unrestricted versions to these tiny groups of customers “until our trusted entree programme is available,” hinting astatine aboriginal plans to grow entree adjacent more. Since the Mythos motorboat successful April, Anthropic has repeatedly emphasized that yet its competitors successful some the backstage and adjacent unfastened value spaces volition inevitably besides connection models with Mythos-level capabilities.

The quality for Claude Mythos and different caller AI models to plan hacking tools that tin find and exploit vulnerabilities successful some caller and bequest bundle has forced tech companies and governments astir the satellite to unafraid their bundle defenses earlier AI models of this level are made broadly disposable to attackers. Anthropic archetypal released Mythos to manufacture partners nether a consortium called Project Glasswing, with the thought that this could springiness members a caput commencement successful preparing their ain systems and weighing planetary solutions to the menace earlier a broader release.

Anthropic wrote successful an update astir Project Glasswing past week: “We’re moving arsenic rapidly arsenic we tin to safely merchandise Mythos-level capabilities successful wide access. To bash so, we’ll request highly robust safeguards that forestall the model’s cyber capabilities from being misused—safeguards that we (and, to our knowledge, each different AI developers) person yet to develop.”