Security Alert: Anthropic holds back public release of its latest model to help companies upgrade their cybersecurity

 It is no secret that as tech firms release powerful new AI systems, these models are turning into a cyber­security nightmare and upending the guardrails that protect the internet. That’s because the same models that are so good at helping engineers create new software are also equally capable at aiding hackers probe and find weak points in software and online services. A single AI agent can today scan for vulnerabilities and potentially take advantage of them faster and more persistently than hundreds of human hackers.

Anthropic has now built a model that it claims is so powerful at find­ing security vulnerabilities that it is withholding its release to the public. It is instead making this model, called Claude Mythos Preview, available to a consortium of over 40 large tech firms so that the model can be used to find and patch security vulnerabilities. According to the firm, the new model has already identified “thousands” of bugs and vulnerabilities in popular software programs, including every major operating system and browser.

There is currently an arms race between hackers and the companies racing to defend their systems. Better and more sophisticated AI models will develop and many will fall into the hands of hackers too. The only solution for these firms is to get their hands on these models before the hackers and to patch up their weak spots.