Two weeks in the past, Anthropic’s secretive AI mannequin often known as Claude Mythos was found as a result of unpublished details about it was sitting in a publicly accessible database. Now the corporate is announcing that it’s teaming with the largest corporations on the earth to let that mannequin free to flag potential safety vulnerabilities inside their programs.
The restricted launch of Mythos, dubbed Venture Glasswing, consists of about 40 organizations that can have entry to a preview model of the mannequin that’s supposedly higher than “all however probably the most expert people” at discovering software program vulnerabilities. Launch companions for the undertaking embrace Amazon Internet Providers, Apple, Google, JPMorganChase, Microsoft, and NVIDIA, amongst others. In response to Anthropic, the early returns from the collaboration have been jarring, as the corporate claims to have discovered “hundreds of high-severity vulnerabilities,” together with some in each main working system and net browser.
It’s unsurprising, given these obvious revelations of significant safety flaws, that Anthropic believes the mannequin “might reshape cybersecurity.” Its benchmark checks definitely appear to indicate that, as Mythos Preview persistently outperformed Claude Opus 4.6, together with on the CyberGym test that seeks to determine how nicely AI brokers can detect and reproduce real-world software program vulnerabilities. The anecdotes help it, too. Anthropic says Mythos discovered a bug within the open-source working system OpenBSD that had been there for 27 years and noticed a sequence of vulnerabilities in Linux that might be used to fully hijack a machine.
What’s attention-grabbing is that simply weeks in the past, when Mythos was first found (because of a quite simple safety slip-up, curious how that one wasn’t caught by the all-seeing machine), Anthropic was apparently positioning the mannequin as being so highly effective that it could current unprecedented cybersecurity dangers. The corporate hasn’t completely backed off that notion—it stated that it gained’t make Mythos Preview out there to the general public due to the dangers it poses to facilitate cybersecurity assaults. However to go from conserving it underneath wraps as a result of it’s too highly effective to launch to deploying it throughout important tech infrastructure is a little bit of a leap.
It’s onerous to take away Anthropic’s positioning of Mythos from the lengthy historical past of AI hype cycles, during which these instruments are offered as world-altering (and probably world-destroying) entities, just for them to be incapable of answering what number of occasions the letter “r” seems in strawberry. Method again in 2019, when Elon Musk was nonetheless at OpenAI, the corporate warned that it had developed a text-generation instrument that was too harmful to be made public. A number of months later, it was launched anyway, and the world saved spinning, simply with a bit extra machine-generated nonsense in it.
Anthropic has run a model of this playbook already because it pertains to cybersecurity. When the corporate dropped Claude Opus 4.6, it touted how the mannequin had discovered hundreds of previously unidentified security vulnerabilities that managed to exist undetected within the wild.
AI fashions like Mythos virtually definitely will play a job—probably even a big one—in the way forward for cybersecurity, working each as a instrument for exploitation and safety. It’ll additionally probably have a endless move of labor in entrance of it, as a result of AI fashions like its cousin Claude hold producing vibe-coded outputs filled with flaws. That’s a method to make sure job safety.
