Project Glasswing

What Project Glasswing’s first month tells us about the security gap that matters

We recently wrote about Anthropic’s Claude Mythos Preview and why it represented a meaningful shift in the security landscape. Rather than make Mythos generally available, Anthropic set up Project Glasswing, a restricted initiative giving around 50 partner organisations access to Mythos for defensive security work. Since then, Anthropic and many of the participants have begun to share the results of Glasswing. The outcomes to date are worth looking at carefully.

What the data shows

Mythos scanned over 1,000 open-source projects and flagged 23,019 vulnerabilities. Of those, 6,202 were estimated as high or critical severity. When 1,900 of the results were reviewed by external security firms, 1,726 were confirmed as valid. That’s a true positive rate of around 90%.

The individual partner findings are equally striking. Cloudflare reported finding 2,000 bugs, including 400 of high or critical severity. Mozilla fixed 271 vulnerabilities in a single Firefox release, ten times the number identified using an earlier Claude model. Microsoft has said its patch releases will “continue trending larger for some time,” attributing this in part to bugs discovered with Mythos.

Some of the specific findings confirmed just how long some of these vulnerabilities had been hiding. Mythos autonomously uncovered a 27-year-old vulnerability in OpenBSD and a 16-year-old flaw in FFmpeg that five million automated test runs had missed. One wolfSSL vulnerability it found would have allowed attackers to forge TLS certificates, making a fraudulent site appear entirely legitimate to an end user.

 

Glasswing isn’t the only initiative

In our previous post, we said that other frontier model providers would follow with equivalent tools, and some have now done so. OpenAI has launched Codex Security, an agentic application security tool designed to analyse a code repository, generate a custom threat model, and identify vulnerabilities with proposed fixes. It has since expanded this into Daybreak, a broader cybersecurity initiative that repositions Codex Security as an enterprise platform with threat modelling, patch validation, and dependency risk analysis built into the development loop. Partners include CrowdStrike, SentinelOne, Cloudflare, Akamai, and Snyk, spanning the full chain from vulnerability discovery through to edge protection and supply chain defence.

The direction across the industry is now consistent: frontier model capability for vulnerability detection is moving from restricted research into integrated tooling, and the window in which this belongs only to a handful of trusted partners is closing.

 

The gap that matters

The finding that matters most from Glasswing isn’t the volume of vulnerabilities found. It’s what happened next. To date, less than 1% of the vulnerabilities Mythos identified have actually been patched. Partners’ rate of bug discovery increased by more than a factor of ten. The capacity to fix those bugs remains the key gap.

Shipping a remediation requires triage, coordination, testing, and deployment across live systems. That process hasn’t accelerated and the result is a widening gap between what can be found and what can actually be closed. That gap is where exposure sits.

The scarce resource in security is no longer the ability to find bugs. It’s the ability to prove which findings are real, decide which ones matter, coordinate disclosure, ship safe patches, deploy them into environments, and verify that the risk is actually gone.

 

What most organisations are missing

Glasswing and Daybreak were built for large enterprises and critical infrastructure providers, however the underlying challenge isn’t restricted to that tier: frontier model vulnerability detection is only as useful as the ability to act on what it finds.

At the same time, AI is producing code at a velocity that traditional security tooling can’t match. The same capability that accelerates development can now find complex, chained vulnerabilities in that code at comparable speed. That makes this a software lifecycle problem, not a security team problem in isolation.

The tools exist, so the question is whether organisations have the context, capacity, and engineering expertise to turn findings into fixed vulnerabilities, and integrate continuously without it becoming a full-time distraction from everything else your team is trying to deliver.

What you should do now

In order to prepare for this new era of AI powered cyber security, organisations need to act urgently. Our recommendation is to start using existing AI-powered tooling now, running code-based vulnerability testing and remediation before that window closes. This will position you meaningfully ahead of those that wait for a formal programme, a scheduled audit or your next penetration test.

The priority is to start finding what’s there. The harder work is building the capacity to validate, prioritise and fix at the same pace you can now discover. This is where most teams will need support and that’s what Atomic is actively helping organisations work through, both in running the tooling and in building the internal capability to act on what it surfaces.

If you’d like to talk through what closing that gap looks like for your organisation, get in touch.

 

Latest from Atomic