Anthropic's AI is finding Microsoft bugs faster than engineers can fix them

Key points

• Anthropic gave select software companies access to a new AI model called Mythos in mid-May 2025 to find security flaws before hackers could.
• Microsoft launched an internal effort, codenamed Project Glasswing, specifically to fix the vulnerabilities Mythos uncovered.
• The effort is framed as a defensive race: patch the holes before China or other adversarial actors use similar AI tools to exploit them.
• Engineers at Microsoft's Redmond headquarters openly questioned whether Mythos performed as well as Anthropic claimed.

Dozen of Microsoft engineers crowded into a conference room in Redmond, Washington, or joined by video, one afternoon in mid-May. The meeting had an unusual agenda: an AI had been breaking their own software, and they needed to talk about how badly.

The AI in question is Mythos, a model built by Anthropic, the San Francisco safety-focused AI company behind the Claude family of chatbots. Anthropic gave Mythos access to a small group of organisations that build software used by ordinary people, businesses and governments. The idea was straightforward: let an advanced AI hunt for security vulnerabilities, the weak spots in code that attackers can slip through, before real-world hackers find them first.

Microsoft's internal response got a name: Project Glasswing.

Why does this matter to ordinary people?

Microsoft software sits on hundreds of millions of computers worldwide. A vulnerability left unpatched is an open door. If a government-backed hacking group, or a criminal gang, finds the same flaws first, the consequences range from stolen data to disrupted infrastructure.

That is exactly the threat Anthropic and Microsoft are trying to get ahead of. As reported by Ars Technica, the concern inside the Glasswing meetings was explicit: adversarial governments, China named among them, could soon deploy their own AI tools to scan for the same weaknesses at scale.

AI models have long been used to help write code. What is newer, and what makes Mythos notable, is the speed at which it apparently identifies flaws in existing code. Fixing software vulnerabilities has always been a race against discovery. Mythos seems to be shortening the clock on both sides.

Were the engineers convinced?

Not entirely, at first. As the Glasswing meeting got underway, one engineer asked the question sitting in the room: did Mythos actually "live up to the hype that Anthropic claimed it would have had?"

That scepticism is healthy. AI systems regularly get oversold. A model that finds ten minor, already-known issues looks different from one that surfaces genuinely dangerous, previously unknown flaws. The source reporting does not yet say which category Mythos's findings fall into, or how many vulnerabilities were found and patched.

What happens next?

The details of what Mythos found remain unpublished, which is standard practice: you do not announce exactly where the holes were until the patches are live.

What this story signals more broadly is a shift in how the security industry works. Companies are starting to use AI offensively in a controlled way, pointing it at their own products, to beat genuine attackers to the punch. That approach will spread. The question is whether defenders can keep the pace.

For everyday users, the practical advice has not changed: keep your software updated. When patches arrive, they now may owe their existence to an AI that found the problem first.

Common questions

What is a software vulnerability, exactly?

A vulnerability is a mistake or weak spot in a program's code that an attacker can use to get in where they should not. Think of it as an unlocked window in an otherwise locked building.

Should I be worried that AI can now find these flaws so quickly?

The same speed that worries defenders could protect you. If companies use AI to find and fix flaws before attackers do, your software becomes more secure. The risk is that hostile actors get the same tools and move faster than the defenders.