Anthropic has built its most capable AI model, and it is keeping it from the public. During testing, Mythos Preview proved unusually good at finding and exploiting serious flaws in software used around the world. The company presents the choice as a security decision.
Anthropic’s technical write-up says the model found and exploited previously unknown security flaws in every major operating system and every major web browser. Many of those flaws were ten or twenty years old, and the oldest confirmed case was a 27-year-old bug in the OpenBSD operating system. In one example, the model combined several flaws to break out of browser and operating system protections that were meant to contain the damage.
Mythos Preview also built working attack code from what it found, the company says. Engineers with no formal security training asked the model to search overnight for flaws that would let an attacker run code on another system. They woke up to a complete working attack. Researchers also built setups in which the model turned flaws into attacks with no human help after the first instruction.
Speed is the business issue. If a model can find a serious flaw and turn it into attack code faster, the window between discovery and misuse shrinks. The most advanced AI, already improving how software gets written, is also getting far better at finding weak points in it and working out ways through them.
That helps explain why the model went to Project Glasswing instead of the public. Glasswing is a cybersecurity coalition that, according to Anthropic, involves Apple, Google, Microsoft, Amazon, NVIDIA and the Linux Foundation. These organisations are getting early access so defenders can secure important systems first. The company is committing up to $100 million in credits for using the model and $4 million in donations to open-source security groups.
Once an AI system can outperform many human experts at finding flaws and building attacks, who gets access to it becomes a question of control. By the company’s own count, Mythos Preview has already found thousands of high-severity flaws. Anthropic says it chose restricted access because capabilities of this kind could spread quickly beyond groups committed to using them safely.
