Project Glasswing is expanding from an initial set of roughly 50 partners to about 150 new organizations. Anthropic says the work is focused on securing important software, including organizations that provide critical infrastructure across more than 15 countries.
The key number
The initial partner activity reportedly found more than 10,000 high- or critical-severity flaws. That is the operational story. Frontier models can increase vulnerability discovery volume, but security teams still need triage, exploitability analysis, patch sequencing, and regression testing.
Anthropic also points to Claude Security, using public frontier models such as Claude Opus 4.8, to scan codebases and suggest patches. That positions model-assisted security as a repeatable workflow rather than a one-off research demo.
How to absorb the signal
Teams should create a separate queue for model-generated security findings. The queue needs ownership labels, affected package data, reproduction steps, severity rationale, patch candidates, and validation evidence. Without that structure, thousands of findings can become noise.
The patch side is equally important. AI-suggested fixes should go through normal review and test gates, especially in memory safety, authentication, parser, and network boundary code. High severity is not the same as high confidence.
Builder takeaway
The next bottleneck in AI security scanning is not discovery. It is disciplined remediation throughput.