Project Glasswing: securing critical software with Claude Mythos Preview
2026-04-07 · Publisher : Anthropic · Domain : Artificial Intelligence
Anthropic launches Project Glasswing, a cyber defense program giving vendors and maintainers of critical software restricted access to Claude Mythos Preview. Within seven weeks, more than 10,000 serious flaws are found.
What changes
- Founding partners: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks.
- Anthropic commits 100 million USD in usage credits and 4 million USD in donations to open-source security.
- By May 22, more than 10,000 high- or critical-severity vulnerabilities identified by about fifty partners.
- On June 2, the program expands to about 150 additional organizations in more than 15 countries (power, water, healthcare).
- The bottleneck shifts from finding flaws to verifying and patching them.
On April 7, 2026, Anthropic announces Project Glasswing, a collaborative program to secure the world's most critical software. It relies on Claude Mythos Preview, a model whose offensive capabilities are deemed too sensitive for public release (see the "Claude Mythos Preview" entry of the same day). This entry covers the program itself. ### What happened - Founding partners are AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, joined by about forty organizations that maintain critical software. - Anthropic commits 100 million USD in Mythos Preview usage credits, 2.5 million USD to Alpha-Omega and OpenSSF (via the Linux Foundation) and 1.5 million USD to the Apache Software Foundation. - Examples cited at launch: a 27-year-old flaw in OpenBSD and a 16-year-old bug in FFmpeg, missed after five million automated test runs. ### Key figures (May 22, 2026 update, Anthropic data) | Metric | Value | |---|---| | High- or critical-severity vulnerabilities found by partners | over 10,000 | | Open-source projects scanned by Anthropic | over 1,000 | | Estimated high or critical flaws in those projects | 6,202 | | True positive rate (1,752 flaws assessed by third parties) | 90.6% | | Flaws disclosed to maintainers / patches shipped | 530 / 75 | | Flaws found by Mozilla in Firefox | 271 | Cloudflare reports 2,000 bugs, 400 of them high or critical severity. ### Expansion On June 2, 2026, Anthropic opens the program to about 150 additional organizations in more than 15 countries, from underrepresented sectors: power, water, healthcare, communications, hardware. ### Caveats According to Anthropic, "finding them in the first place has become vastly more straightforward": the difficulty shifts to verification and patching. A patch takes two weeks on average to deploy, and some maintainers have asked for a slower disclosure pace. ### Key takeaways Glasswing introduces a restricted, defensive access model for offensive-capable models. It anticipates the arrival of cheap models with comparable cyber capabilities.
Links
Tags : IA, Anthropic, Cybersécurité, Mythos, Open Source, Accès restreint