Anthropic's zero‑day finder

A zero-day is a software bug that nobody has publicly documented yet, which means defenders have had zero days to prepare a patch before an attacker could use it. Anthropic says its new model, Claude Mythos Preview, can find these hidden flaws in real codebases and is already being used in private security work. (red.anthropic.com) Anthropic wrapped that effort in a program called Project Glasswing, which it says is aimed at “the world’s most critical software.” The launch partners Anthropic lists are Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. (anthropic.com) The model is not being pitched as a chatbot for general coding help. Anthropic’s own writeup says Mythos Preview is being used to scan software, find previously unknown vulnerabilities, and help defenders fix them before details are published. (anthropic.com) Anthropic also says more than 99 percent of the vulnerabilities it has found are still unpatched, which is why it is withholding technical details for now. That is the logic of coordinated vulnerability disclosure: tell the vendor first, give them time to fix it, then disclose enough for everyone else to protect themselves. (red.anthropic.com) (cisa.gov) That workflow already exists in security, but it has usually been driven by human researchers or dedicated teams like Google Project Zero. Google’s disclosure policy gives vendors a fixed clock, and its 2025 policy update described the current model as “90+30,” meaning 90 days for a fix and 30 more before full public disclosure when conditions are met. (googleprojectzero.blogspot.com) What changes here is speed and scale. Anthropic says Mythos Preview can find and exploit zero-days in open-source software, reverse-engineer exploits for closed-source software, and even turn some known-but-unpatched flaws into working exploits. (red.anthropic.com) Anthropic had already warned in February 2026 that large language models were getting meaningfully better at discovering zero-days. The new announcement is the next step after that warning: instead of only measuring the risk, Anthropic is now putting the capability inside a private remediation pipeline with large vendors. (red.anthropic.com) (anthropic.com) That private pipeline matters because the dangerous part of a zero-day is not just finding it but the patch gap after discovery. Google Project Zero reported that vendors averaged 52 days to fix vulnerabilities it reported in 2021, and CISA keeps a Known Exploited Vulnerabilities catalog precisely because many flaws are weaponized before organizations patch them. (googleprojectzero.blogspot.com) (cisa.gov) Anthropic says over 40 additional organizations beyond the named launch partners have also been given access to use the model on first-party and open-source systems. That means the company is trying to place the tool where widely used infrastructure is maintained, not just where consumer apps are built. (anthropic.com) The hard part is that a model that helps defenders can also help attackers if it leaks or is widely released. Anthropic’s system card for Claude Mythos Preview says the model shows a “striking leap” in cybersecurity capability, which is why the company is pairing the release with restricted access and detailed safety evaluations. (anthropic.com) So this is not just a product demo from April 2026. It is one of the clearest signs yet that frontier artificial intelligence labs are moving from talking about offensive cyber capability in theory to building real-world systems for private bug hunting, triage, and repair before the public ever sees the bug report. (anthropic.com) (red.anthropic.com)