Claude Mythos isn’t going public — it’s being used to fix some very old internet problems
Anthropic’s Project Glasswing has quietly turned into a pretty serious industry effort.
What started as a cautious rollout is now a full coalition, with AWS, Google, Microsoft, Apple, NVIDIA, Cisco, CrowdStrike, Palo Alto Networks, JPMorganChase and the Linux Foundation all involved – and all working with Claude Mythos behind closed doors.
The model itself is staying locked up, and it’s not hard to see why.
New benchmarks show a big jump from Anthropic’s previous model, especially when it comes to coding on its own.
In testing, Mythos uncovered high-severity bugs that had been sitting around for decades – including a 27-year-old issue in OpenBSD and a long-missed flaw in FFmpeg.
It also managed to chain together smaller vulnerabilities to take full control of a Linux system.
That’s where things get a bit tricky. It’s great at finding and fixing problems, but it can also exploit them – which is why Anthropic isn’t putting it out into the wild.
Glasswing is basically about getting ahead of that risk. The idea is to patch as much as possible now, before similar tools inevitably show up elsewhere.
To help with that, Anthropic is offering $100 million in usage credits for critical infrastructure groups, plus $4 million in funding for open-source security teams.
There’s also some tension building on the government side. The US Department of Defense has labelled Anthropic a “supply chain risk” after the company refused to allow unrestricted military use of the model, particularly around autonomous weapons and mass surveillance. That situation is now playing out in court.
For everyone else involved, access doesn’t come cheap either, with pricing set well above standard AI models — very much an expert-level tool.
Overall, this feels less like a typical AI launch and more like damage control — just at a much bigger scale.
