Anthropic’s Project Glasswing

Software security usually fails in the boring places: one unchecked input, one old library, one access-control rule nobody noticed in a codebase with millions of lines. Anthropic says its new Claude Mythos Preview model can now find and exploit those flaws at a level beyond all but the most skilled human researchers. (anthropic.com) Most automated security tools work like metal detectors at an airport. They look for known patterns, which catches common mistakes, but Anthropic says newer model-based systems can read code the way a human reviewer does and follow how data moves across an application. (anthropic.com) That difference matters because the hardest bugs are not obvious syntax mistakes. Anthropic says rule-based scanning often misses business-logic flaws and broken access control, which depend on understanding how separate parts of a program interact. (anthropic.com) Anthropic’s claim is that Mythos Preview crossed a line from “helpful coding assistant” to “serious vulnerability hunter.” In its April 7, 2026 technical write-up, the company said the model could identify and exploit zero-day vulnerabilities, meaning previously unknown bugs, in every major operating system and every major web browser during testing. (red.anthropic.com) The company also said many of the bugs it found had been sitting in software for ten or twenty years. Anthropic gave one example of a now-patched 27-year-old bug in OpenBSD, which is an operating system with a long reputation for security hardening. (red.anthropic.com) That is why Anthropic did not put Mythos Preview on the open market. In its system card, the company said the model showed a large jump over Claude Opus 4.6 and that the capability increase led Anthropic to keep it out of general release and limit it to a defensive cybersecurity program. (anthropic.com) Project Glasswing is the container Anthropic built around that decision. The program launched on April 7, 2026 with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks as launch partners using the model for defensive security work. (anthropic.com) Anthropic says more than 40 additional organizations that build or maintain critical software infrastructure also got access to scan both their own systems and open-source software. The company paired that with up to $100 million in Mythos Preview usage credits and $4 million in donations to open-source security groups. (anthropic.com) The pitch is simple: let defenders use the stronger lockpick before attackers get one. Anthropic says Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and browser, and argues that the spread of similar capabilities is now a matter of months, not years. (anthropic.com) Anthropic is also trying to set rules for what happens after the model finds a bug. Its disclosure policy, updated March 6, 2026, says the company aims to notify vendors quickly, provide human-reviewed reports with suggested fixes, and generally follow a 90-day disclosure timeline or wait until a patch ships, whichever comes first. (anthropic.com) The unresolved part is scale. Anthropic says more than 99% of the vulnerabilities Mythos Preview found are still unpatched, which is why it is withholding details, and that number hints at the next problem: even if artificial intelligence can find bugs faster, software vendors still have to triage, patch, test, and deploy those fixes one by one. (red.anthropic.com)