Project Glasswing just produced its first confirmed result.
Daniel Stenberg, the lead developer of curl and libcurl, was contacted by representatives of the Linux Foundation’s Alpha Omega project and offered access to Anthropic’s Mythos model, the same restricted-capability AI that’s been under governance scrutiny since Anthropic disclosed its existence in April. Stenberg accepted, signed a contract, and ran the model against the curl codebase. According to multiple independent reports, Mythos identified a vulnerability reportedly approximately 27 years old.
That’s a long time for a bug to survive.
The vulnerability’s classification hasn’t been confirmed in publicly available source text, but reporting describes it as an integer overflow type, reportedly classified as CWE-190, pending confirmation from Stenberg’s full post. The “178K lines of code in one session” figure circulating in some coverage has no verifiable source and doesn’t appear in this brief.
Project Glasswing Access Chain
The mechanism matters here. Alpha Omega, the Linux Foundation initiative funded by Microsoft and Google to improve security across critical open-source software, brokered Stenberg’s access to Mythos under what’s described as the Project Glasswing arrangement. That’s not Anthropic going directly to an open-source maintainer. There’s an intermediary governance layer: the Linux Foundation holds the access agreement, Stenberg signed a contract, and Mythos ran under that structure. Whether that structure adequately addresses disclosure timelines and liability for what Mythos finds is a different question, and one that hasn’t been answered publicly yet.
The catch is that Glasswing’s output is only useful if the downstream patch and disclosure process is equally well-governed. A frontier AI finding a bug is the easy part. Coordinating a responsible disclosure across a library used in millions of deployments, that’s the hard part, and no public details are available yet on how that process is being handled for this specific finding.
For enterprise security teams, the signal is clear: curl and libcurl are foundational to HTTP transfers across the software stack. A previously unknown integer overflow vulnerability in a library this widely deployed has real exposure implications, even before patch availability is confirmed. Watch Stenberg’s blog and the curl security advisory tracker for disclosure details.
For compliance professionals tracking capability assessment: this event directly feeds the CAISI framework discussions covered in the May White House mandatory AI model review brief. Mythos finding a real-world vulnerability in critical infrastructure is the type of confirmed capability output that regulators building capability thresholds will reference. It’s evidence, not just a vendor claim.
What to Watch
Anthropic has long maintained that Mythos represents a restricted-access model with offensive cybersecurity capabilities significant enough to require special governance. The April coverage on who has access to Mythos documented the structure of that restriction. This is the first time the restriction has produced a publicly confirmed, independently verifiable output, Stenberg’s post, with his name on it, is as close to ground truth as this story gets.
Wait for the full haxx.se post and the curl security advisory before making any patch decisions. The vulnerability age and classification are still reported, not confirmed in the primary source text.