Project Glasswing is a program by the AI company Anthropic, started in April 2026. Companies and organisations that maintain widely used software, such as Microsoft, Apple and the Linux Foundation, get early access to Anthropic's most capable AI model to find and fix security holes before attackers find them. For a small business the effect is indirect: the fixes arrive as updates, and Anthropic's advice is to install them quickly.
How does it work?
Partners point the model at their own code. The model reads the code, looks for flaws and can write an exploit, a working attack that proves a flaw can be abused. At the launch on 7 April 2026, Anthropic said the model, Claude Mythos Preview, had found nearly all of the flaws it described, and built many of the exploits, without human steering.
People then check each finding. According to Anthropic's first update (22 May 2026), Anthropic or an independent security firm reproduces the flaw, reassesses how severe it is, checks whether a fix already exists and writes a report for the software's maintainers, who build and release the patch. Details stay private until users have had time to update: the industry convention is 90 days after discovery, or about 45 days after a patch appears.
Who takes part?
Anthropic started with eleven launch partners, among them Amazon Web Services, Apple, Cisco, Google, JPMorganChase, Microsoft and the Linux Foundation, plus more than 40 other organisations that build or maintain critical software. It committed up to $100 million in usage credits and $4 million in donations to open-source security organisations. On 2 June 2026 it added about 150 organisations in more than 15 countries, from sectors such as power, water, healthcare, communications and hardware.
What has it found?
At the launch, Anthropic reported thousands of previously unknown flaws (zero-days), including a 27-year-old one in OpenBSD that let an attacker crash any machine running it just by connecting to it. In the update of 22 May 2026, Anthropic and about 50 partners had found more than ten thousand high- or critical-severity vulnerabilities in the first month. In a separate scan of more than 1,000 open-source projects, 1,752 findings that the model rated high or critical were checked, mostly by independent firms: 90.6% were real flaws, and 62.4% were confirmed as high or critical. The post what Project Glasswing found in its first month has the details.
Where is the bottleneck?
Finding flaws is no longer the slow step. Anthropic wrote that progress is now "limited by how quickly we can verify, disclose, and patch the large numbers of vulnerabilities found by AI". By 22 May it had reported an estimated 530 high- or critical-severity bugs to open-source maintainers, and 75 of them were patched. Maintainers were already facing "a deluge of low-quality, AI-generated bug reports", and some asked Anthropic to slow down.
Why only selected partners?
Anthropic expects other AI companies to build models with similar skills, possibly without safeguards against misuse, and wants the defenders of the most important software to get a head start. In May 2026 it wrote that no company, itself included, had safeguards strong enough to prevent such models from being misused. On 9 June 2026 Glasswing partners could switch to Claude Mythos 5, released at the same time as Claude Fable 5: the same model with cyber safeguards, for general use.