Summary: Anthropic launched Project Glasswing on April 7, 2026, giving partner organizations access to the unreleased Mythos Preview model for defensive vulnerability discovery, black-box testing, endpoint security, and penetration testing. The announcement reported CyberGym 83.1% versus Opus 4.6’s 66.6%, thousands of pre-launch zero-day findings, up to $100 million in credits, and planned 90-day follow-up reporting.
Anthropic chose April 7, 2026 to put an unusual project on the record: Project Glasswing, whose stated mission is to keep critical software safe in the AI era. The founding roster spans AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, with access also going to more than 40 other organizations whose work builds or maintains critical software infrastructure.
The alliance, however, is not the headline. The headline is what every member is handed: Claude Mythos Preview, a frontier model the public has never seen. Anthropic is blunt about what it can do — at finding and exploiting software vulnerabilities, the model “can surpass all but the most skilled humans.” Once offensive skill reaches that point, defending software stops being any single company’s burden.
A Model Built to Dig Out Vulnerabilities on Its Own
Under Glasswing, partners aim the model at the defensive side of the ledger: local vulnerability detection, black-box testing of binaries, endpoint security, and penetration testing. The benchmark record supports that framing — Mythos Preview reached 83.1% on the CyberGym security evaluation, where the previous-generation Opus 4.6 managed 66.6%. In its April 7 write-up, TechCrunch described the program as the most ambitious cybersecurity initiative Anthropic has attempted.
There is wordplay in the name as well. The glasswing butterfly (Greta oto) flies on wings you can see straight through — a reminder that these bugs were never actually hidden, and that seeing clearly is itself a form of defense.
A Dozen Founding Partners, Backed by $100 Million in Credits
Nobody would mistake this for a signature-only coalition. Anthropic committed real resources: Mythos Preview usage credits worth as much as $100 million, plus $4 million donated to open-source security, split between $2.5 million for Alpha-Omega/OpenSSF and $1.5 million for the Apache Software Foundation. Fortune’s coverage flagged the strangest part of the roster — Apple, CrowdStrike, and Google spend most days competing with one another, yet here they share one defensive list.
Pricing came out at the same time: once the preview ends, the model costs $25 per million input tokens and $125 per million output tokens, delivered through the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. The takeaway: this security capability reaches customers over the ordinary enterprise cloud channels.
What Turned Up Before Launch Day
The model arrived with results already on file. Ahead of the announcement it had surfaced thousands of zero-day vulnerabilities, and no major operating system or web browser came away untouched. Three findings convey the scale: a remote-crash bug that had sat inside OpenBSD for 27 years; a 16-year-old FFmpeg defect that automated tests ran into five million times without ever flagging it; and a Linux kernel exploit chain that climbs from user access to full control of the machine.
Within 90 days, Anthropic has pledged to publish a report gathering the lessons learned and the vulnerabilities now fixed, with best-practice guidance on disclosure, patching automation, and supply-chain security to follow.
Why the Model Stays Off the Open Market
General users will not get their hands on Mythos Preview. The logic is plain: sooner or later, capabilities of this kind reach actors with no intention of deploying them safely, against a backdrop where cybercrime already burns roughly $500 billion a year. Guardrails that block the model’s most dangerous outputs are slated to arrive with a future Opus-class model.
Over the medium term, coordination moves to an independent third-party body, so public-private defense collaboration no longer rests on any single model company — a direct answer to the familiar worry about a lab being both referee and player.
The Reading for Security and Dev Teams
Three consequences stand out. Defenders get a window, and it closes: run this caliber of bug-hunting across your own estate before adversaries obtain anything comparable. The bottleneck then shifts from discovery to remediation — once flaws arrive by the thousand, disclosure, triage, and patching must be automated, or every finding is just more inventory. And open-source infrastructure takes the hardest hits, which is why the $4 million in donations flows to OpenSSF and Apache, the stewards of the projects where the real patching backlog lives. Product teams should hear the same message: wiring security scanning and repair pipelines into agent workflows is a line item that can no longer be put off.