Anthropic Offers Mythos Upgrade for Cyber ​​Partners and ‘Safe’ Version for All.

Anthropic released two new AI models called Claude Fable 5 and Claude Mythos 5 on Tuesday, which the company says have greater potential than Myth Take a look at the overview was released in April to a small group of technical staff. Anthropic has said that the initial, limited release was based on concerns that the model’s capabilities could be used by bad actors to create hacking tools that could catch on. protectors don’t care.

Anthropic is currently only releasing Claude Mythos 5 to a small group of users, many of whom received access to the Mythos Preview, and the company says it is cooperating with the US government on the release.

Claude Fable 5, which is being released publicly, uses the same model as Mythos 5, but it will have “supervisors” at launch, the company said on Tuesday, which will prevent the model from answering many questions related to cybersecurity, biology, and chemistry. These requests will be sent to the previous version of AI, Claude Opus 4.8. If Anthropic suspects the user is trying to do it distillation-training a small AI model for larger AI-based solutions-on Claude Fable 5, the requests will also be sent to Claude Opus 4.8, the company says.

In an interview with WIRED, Anthropic’s head of product management, Diane Penn, says the company has been grappling with the question of how to handle Mythos’s vulnerability-discovery software and other advanced capabilities since before April’s release, but testing and user input since then helped pave the way.

“We’re trying to improve in a productive way, even though we don’t have the perfect (solution) for every problem,” says Penn. “In a lot of ways, this seemed like the most feasible and the best. We just felt like this was the best way to get users to get the most out of Fable 5.”

Currently, Penn says the defense system is designed to be scalable, meaning that some user queries can be sent to a slower AI version even if they are not performing well. Over time, Anthropic hopes to make its teams more accurate, but Penn says this is the only safe way the company can release the product at this time.

The company said on Tuesday that in addition to offering Claude Mythos 5 to Project Glasswing partners, it is also offering a selection of “biological researchers.” In addition, Anthropic became known in its own blog post of Tuesday’s launch that it is offering unlimited models to a small group of customers “until our reliable access program is available,” and hinted at future plans to expand access. Since the launch of Mythos in April, Anthropic has repeatedly emphasized that competitors in the private and open space will also offer models with Mythos capabilities.

The ability of Claude Mythos and other new types of AI to create hacking tools that can find and exploit weaknesses in both new and popular software has forced technology companies and governments around the world to protect the security of their software before these types of AI are released. Anthropic first released Mythos to its corporate partners under a contract called Project Glasswing, with the idea that this would give members a head start in planning their systems and evaluating global responses to threats before release.

Anthropic wrote in change it about Project Glasswing last week: “We’re working as fast as we can to intelligently release Mythos-level capabilities into the mainstream. To do that, we’ll need very strong defenses that will prevent the model’s cyber capabilities from being misused—defenses that we (and, to our knowledge, all other AI developers) have yet to develop.”