Anthropic says roughly 50 Project Glasswing partners have already found more than 10,000 high- or critical-severity vulnerabilities with Claude Mythos Preview.
What Project Glasswing Is Signaling
Anthropic’s Project Glasswing pairs Claude Mythos Preview with roughly 50 partners whose job is simple on paper and hard in practice: find real security defects before attackers do. The headline result is already large: more than 10,000 high- or critical-severity vulnerabilities. That number is less interesting as a trophy and more useful as evidence that AI-assisted review can surface serious issues at a volume traditional manual cycles rarely match when coverage is broad and time is short.
High- and critical-severity findings matter because they usually map to concrete failure modes—broken access control, injection paths, unsafe deserialization, credential exposure, or privilege boundaries that do not hold. When a coordinated program produces thousands of those, the lesson is not “AI replaces security teams.” It is that model-assisted analysis can expand the surface area of review faster than many organizations currently staff for, especially across large codebases, mixed stacks, and services that rarely get a full adversarial pass.
Why Partner Programs Amplify Results
A preview model alone does not create 10,000 findings. Partners do. They bring domain context, product architecture knowledge, threat models, and the judgment to decide which model outputs are real issues versus noise. Glasswing’s design—many partners, one strong model surface—spreads specialized expertise across different codebases while keeping the analysis tooling consistent enough to compare patterns.
That structure has practical tradeoffs. Breadth finds more issues; consistency in severity labels becomes harder. Without shared triage rules, “critical” can mean different things to different teams. Programs that want durable value need agreed severity criteria, clear reproduction requirements, and a path from model suggestion to verified defect. The partner model works when humans stay accountable for validation and prioritization, not when every generated alert is treated as an incident.
How Security Teams Should Use a Signal Like This
Treat the Glasswing result as a capacity argument, not a shopping list of CVEs. If AI-assisted review can uncover large volumes of severe issues in partner environments, your own backlog likely contains similar latent risk in places you have not fully instrumented: legacy services, admin tooling, internal APIs, and integrations that sit outside the usual penetration-test scope.
- Define which systems are in scope for AI-assisted review and which must stay offline or heavily sandboxed.
- Require reproducible evidence for every high- or critical-severity claim before ticket creation.
- Route findings into existing severity and SLA frameworks so AI output does not create a parallel shadow queue.
- Measure false-positive rate and time-to-confirm; those metrics decide whether the program scales or stalls.
Also plan for disclosure and fix velocity. Finding volume only helps if engineering can patch, ship, and verify. Pair model-driven discovery with ownership maps, patch SLAs for high/critical issues, and regression tests that lock down the fixed path. Otherwise the program becomes a report generator instead of a risk reducer.
Practical Guardrails When Adopting Similar Workflows
Using a capable preview model on production-adjacent code raises its own risks: secret leakage into prompts, over-trust in fluent but wrong explanations, and scope creep into systems that should not leave a controlled environment. Keep analysis inside approved tooling boundaries, strip credentials from inputs, and log what was sent and returned so audits can reconstruct decisions later.
Start narrow: one high-value service, clear severity definitions, dual review for criticals, and a weekly reject-rate review so the team learns where the model is strong and where it hallucinates. Scale only after triage quality is stable. Project Glasswing’s early tally—tens of thousands of severe findings across a few dozen partners—shows the upside of structured AI-assisted hunting. Capturing that upside in your org depends less on the model name and more on verification discipline, fix capacity, and honest severity control.