Anthropic seems to be making ready for the overall availability of Mythos, which was introduced in April as a restricted mannequin that poses vital safety dangers to non-public and public software program.
On April 7, Anthropic introduced an early preview of Mythos, calling it a brand new frontier mannequin with surprisingly superior options for pc safety duties.
In accordance with Anthropic, the Mythos mannequin gives vital enhancements in code reasoning and autonomy, far exceeding its present flagship mannequin, Opus 4.7.
Bettering coding is nothing new for AI fashions, however within the case of Mythos, Anthropic discovered that the mannequin was in a position to robotically develop practical cyberattacks at a extremely specialised degree.
The corporate additionally claimed that its deployment poses critical dangers to the world’s digital infrastructure.
“Those that can benefit from these instruments can have a bonus,” Antropic warned.
“If Frontier Labs isn’t cautious with how these fashions are launched, within the quick time period it might turn out to be the attackers. In the long run, we anticipate it to be the defenders who handle assets extra effectively and use these fashions to repair bugs earlier than new code is launched.”
To forestall attackers from exploiting the massive variety of unpatched vulnerabilities in common apps like Firefox, Anthropic has determined to not make the Mythos mannequin publicly accessible till it has a robust guardrail system in place.
With Claude Code and Claude Safety referencing the Mythos mannequin, it appears Anthropic could have developed a robust guardrail system.
The truth is, some customers briefly observed a toggle to allow Mythos within the public model of the Claude code earlier than it was taken offline.
This mannequin, known as claude-mythos-1-preview, additionally briefly appeared within the public model of Claude Safety.
This confirms that Anthropic is making ready to make this mannequin accessible to the general public, nevertheless it’s unclear if it is going to be accessible to all subscription tiers.
Anthropic mentioned it’s working with firms to uncover potential AI-powered exploits.
Anthropic has confirmed that it’s engaged on a brand new venture known as “Glasswing.” On this venture, AI startups collaborate with different firms to guard the world’s most important software program from potential AI exploits.
This effort makes use of the unreleased Claude Mythos Preview and has already been in a position to help as much as 50 organizational companions.
The Mythos mannequin was in a position to uncover 10,000 high-severity or high-severity vulnerabilities within the first month alone. This explains why Anthropic has been holding off on releasing it to the general public.
Anthorpic at the moment gives Claude Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, and Haiku 5.5.
Automated penetration testing instruments supply actual worth, however they have been constructed to reply one query: Can an attacker get by means of your community? They aren’t constructed to check whether or not controls block threats, detection guidelines hearth, or cloud configurations are preserved.
This information describes six surfaces that it’s best to truly study.
Obtain now
