What did Project Glasswing accomplish for Anthropic?
Claude Mythos Preview found over 10,000 critical bugs
Anthropic says its Claude Mythos Preview has been used to identify more than 10,000 high- or critical-severity vulnerabilities since the launch of Project Glasswing.
The significance is the scale and severity mix. Finding that many high-impact issues suggests the system wasn’t just generating “weak signals” or low-confidence reports; it was participating in a vulnerability discovery workflow that produced triage-able findings at a level that security teams typically prioritize.
Project Glasswing, as described through Anthropic’s announcement, is framed as an ongoing effort to accelerate software security research using Claude-based capabilities. By tying the results directly to Mythos Preview, Anthropic is effectively positioning the preview model as an operational tool in its coordinated vulnerability-finding process rather than a purely experimental research artifact.
From a risk perspective, large vulnerability-discovery efforts can affect multiple stakeholders at once:
- Software vendors and maintainers get leads that may be turned into patches and advisories.
- Enterprises can use the results to assess exposure in their own environments.
- Security researchers get additional data points on how effective AI-assisted auditing may be.
It also underscores that vulnerability discovery is increasingly linked to model capabilities and workflows. As Anthropic reports results at the “high/critical” level, it suggests that modern AI tooling can contribute meaningfully to real-world security outcomes.
At the same time, the announcement focuses on the count of vulnerabilities and their severity category rather than detailing which products were affected, how the findings were validated, or how many were ultimately fixed.
Still, for the security industry, the headline number is the message: Project Glasswing and Claude Mythos Preview appear to be producing a steady stream of serious vulnerability leads, supporting Anthropic’s broader push to make AI a part of defensive software engineering.