Anthropic Builds Its Strongest Model, Then Gives It Only to Defenders: Inside Project Glasswing
Claude Mythos Preview found thousands of high-severity vulnerabilities, so Anthropic is limiting it to security partners with $100 million in credits.
Key takeaways
- Claude Mythos Preview is an unreleased frontier model that Anthropic says found thousands of high-severity bugs, including in every major OS and browser.
- Access started with 11 named partners plus Anthropic and 40-plus other organisations; one report later put it near 150 groups in 15 countries.
- Anthropic committed up to $100 million in usage credits and said Mythos will not get a standard public release.
On April 7, 2026, Anthropic unveiled Claude Mythos Preview, a frontier model it described as better at finding software vulnerabilities than all but the most skilled humans, and in the same breath said it would not release it publicly. Instead, the model powers Project Glasswing, a controlled programme that hands it to defenders first.
What Anthropic announced
Project Glasswing pairs Anthropic with a set of launch partners: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia and Palo Alto Networks. Anthropic also extended access to more than 40 other organisations that build or maintain critical software infrastructure. Coverage differed on exact counts, with some outlets citing 12 partners and others roughly 50 organisations, which probably reflects different ways of counting the launch group and the wider tier.
To make the programme usable, Anthropic committed up to $100 million in usage credits for the preview. The company said Mythos Preview had already uncovered thousands of high-severity vulnerabilities, including some in every major operating system and web browser. It stressed the model's job: scanning code, confirming real flaws and helping maintainers patch them before attackers get there. Reports described the approach as the opposite of a product launch. There was no public API, no enterprise rollout and no standard commercial release.
Why hold it back
The logic is dual use. A model that can locate and exploit unknown flaws across operating systems is a powerful defensive tool and a dangerous offensive one. By seeding it with vendors and maintainers first, Anthropic aims to shift the balance toward patching before similar capabilities reach less careful hands. It is a different release philosophy from the one governing earlier Claude models, which shipped broadly with safeguards, and it follows from the company's own safety framework that ties deployment to measured capability.
The strategy has costs. A restricted model cannot be independently benchmarked by the public, so claims about its ability rest largely on Anthropic's account and its partners' feedback. Coverage reported no benchmark scores for Mythos Preview. Access control also becomes a security problem of its own: a later report said Anthropic was investigating unauthorised access to the model tied to a vendor environment, and that further Mythos-class models were expected once safeguards were ready.
Why it matters
Mythos reframes the release question for the whole industry. If capability gains in code reasoning keep translating into vulnerability discovery, the labs will face the same choice repeatedly: ship widely and accept misuse, or gate access and accept scrutiny over who decides. Governments and critical-infrastructure operators will want a say, and the partner list, heavy on cloud, chip and finance firms, shows who is currently in the room.
For software maintainers, the immediate effect may be a surge of high-quality bug reports. Open-source projects with few volunteers could struggle to triage them, which is part of why the Linux Foundation is among the launch partners.
The defender's dilemma
Security teams have long worried about the asymmetry between finding and fixing bugs. Attackers need one working flaw; defenders must find all of them and ship patches across millions of deployments. An AI that can audit large codebases quickly could narrow that gap, but only if patch capacity keeps up. Discovery at machine speed produces a backlog that maintainers, often unpaid volunteers, must clear. Part of the value of the credits and partner structure is to pair finders with the organisations that can actually ship fixes.
There is also the question of precedent. Few frontier developers have announced a model and declined to sell it. Anthropic has argued that deployment decisions should follow measured capabilities, and Glasswing is the first large test of that stance in a commercial setting where competitors are shipping freely. How it plays out will shape expectations for regulators weighing mandatory pre-release testing.
What to watch
- Whether Anthropic publishes details on the vulnerabilities fixed and the disclosure timelines.
- The conditions under which a Mythos-class model becomes generally available.
- Rival labs' stance: matching the gated approach or releasing comparable capabilities openly.
- Findings from the reported investigation into unauthorised access.
Sources
- Anthropic Launches Project Glasswing With Claude Mythos, the Model It Won't Release — NYU Shanghai RITS
- 5 Things To Know On Anthropic's Claude Mythos And 'Project Glasswing' — CRN Asia
- Anthropic's Claude Mythos Breakthrough: 'Project Glasswing' Launched to Prevent AI Cyber-Crisis — Outlook Business
- Anthropic Debuts Claude Mythos For Cybersecurity — Let's Data Science
