Anthropic's most powerful AI model just found thousands of critical security flaws across every major operating system and browser, including a bug sitting undetected in software for 27 years. That model isn't being released to the public. Instead, it's being handed to defenders. That's Project Glasswing.
What is Project Glasswing?
Launched on April 7, 2026, Project Glasswing is a cybersecurity initiative built around Claude Mythos Preview, a frontier AI model that spots serious vulnerabilities humans have missed for years, sometimes decades. In early testing, it found thousands of high-severity flaws across every major operating system and web browser.
Anthropic isn't releasing it publicly. Instead, controlled access goes to vetted partners Amazon, Apple, Microsoft, Google, Cisco, NVIDIA, CrowdStrike, JPMorgan Chase, the Linux Foundation, and 40+ other organizations strictly for defense, scanning codebases, running penetration tests, patching weaknesses before attackers find them. Every partner must share findings with the broader industry.
Why is Mythos not for everyone?
A model that can find vulnerabilities can also exploit them. Releasing it openly would give anyone access to a powerful attack tool, including hackers.
This is where harnessing AI capability becomes critical. The AI guardrails are what make it deployable without opening AI systems to attack.
"What Anthropic has done here is practice what AI safety preaches. They built a powerful capability, recognized the risk, and chose controlled deployment over wide release. That's the kind of decision-making the entire industry needs to normalize."
Prashanna Rao, Head of Engineering, GoML
They're also testing new safety filters through an upcoming Claude Opus model refining guardrails on a less risky system before applying them to something as powerful as Mythos. The approach, for now, appears to be to build the safety net first, then expand access.
Glasswing: Threat or solution for AI safety?
Both, depending on who controls it. Anthropic's argument is that this capability was going to exist regardless. The question is whether defenders or attackers get there first.
The implication runs deeper than cybersecurity. As AI models grow more capable, the gap between what they can do and what they should do keeps widening.
Harnessing that gap with AI safety policies, controlled access, and mandatory disclosure bridges capabilities to outcomes with engineering discipline. Skip it, and every improvement in model capability is also an improvement in potential for harm.
"Guardrails are not a product feature, they are an engineering discipline. Project Glasswing stress tests and improves those guardrails in real time. That feedback loop is exactly how you build AI capability you can actually trust."
Prashanna Rao, Head of Engineering, GoML
The initiative also enlists small open-source software teams with zero security budget who maintain code that power hospitals, grids, and logistics networks. Anthropic has committed $4 million to open-source foundations, so solo developers get the same security scanning that only large enterprises could afford before.
The implications of Glasswing
With $100 million in usage credits committed and findings published within 90 days, Project Glasswing shows AI capability and AI safety don't have to conflict.
We’ve always maintained that the key to production ready AI is in how you harness the model and not the model’s inherent guardrails alone. Agnostic of the current state, AI safety will be a persistent discussion. Even if you do not buy the hype and drama, it will pay to dedicate engineering effort to implementing AI safety harnesses in your real-world AI systems. We have covered some of this in our blog and whitepaper on AI system design.
As for Glasswing, every new AI breakthrough brings a choice. Do you release it quickly and deal with the consequences later, or do you put the right controls in place first?
The optimistic take is that Anthropic chose the second path to turn powerful AI from a potential risk into something useful and safe. The cynical take is that Dario Amodei is merely carving out a market position for Claude based on politics and being seen as the anti-Altman.
