Anthropic's defensive AI push

Software breaks in tiny, boring places: one unchecked input field, one memory error, one old library nobody patched. Anthropic says its new model can spot those hidden cracks well enough that it is not releasing the system to the public at all. (anthropic.com) The kind of flaw Anthropic is talking about is a zero-day vulnerability, which means a bug defenders have not seen yet and attackers could use before a patch exists. In its technical write-up, Anthropic says Claude Mythos Preview can find and exploit zero-day bugs in real open-source codebases and can also turn known-but-unpatched flaws into working attacks. (red.anthropic.com) That is why the company paired the model with a gated program instead of a chatbot launch. Project Glasswing gives selected defenders access first, with Anthropic saying the goal is to secure critical software before similar capabilities spread more widely. (anthropic.com) The partner list shows what Anthropic means by “critical.” The launch group includes Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. (anthropic.com) Anthropic says it also extended access beyond those big names to more than 40 additional organizations that build or maintain important software infrastructure. The company says those groups can use the model to scan both their own systems and open-source software that other companies depend on. (anthropic.com) This is not Anthropic’s first warning that artificial intelligence can speed up hacking. In a December 2025 post, the company said a Chinese state-sponsored group used Claude Code in an espionage campaign to inspect systems and identify high-value databases faster than human attackers alone. (anthropic.com) Anthropic has been building the defensive side of that argument for months. In March 2026, it launched Claude Code Security, a product that reads code more like a human security researcher and tries to verify its own findings before showing them to analysts. (anthropic.com) Project Glasswing pushes that logic much further. Anthropic says Mythos Preview is strong enough at computer security tasks that the company does not plan to make it generally available, even while it publishes a system card and shares lessons from the testing. (red.anthropic.com) The bet is simple: if frontier models are going to become excellent bug hunters, the first people using them should be the ones fixing the locks, not the ones picking them. Anthropic says it will share what it learns from Glasswing so the broader software industry can harden systems before this style of attack becomes commonplace. (anthropic.com)