The Short Version
Anthropic published its first Project Glasswing progress report on May 22: in approximately 30 days, Claude Mythos Preview and roughly 50 partner organizations identified more than 10,000 high- or critical-severity vulnerabilities in systemically important software. Separately, The Next Web reported that US intelligence agencies continue actively using Claude despite the federal government designating Anthropic a national security concern over Mythos’s cyber capabilities.
What Actually Happened
Project Glasswing launched on April 7, 2026, as Anthropic’s defensive cybersecurity initiative: deploying Claude Mythos Preview with approximately 50 trusted partner organizations to proactively find vulnerabilities in critical software before adversaries do. The May 22 progress report disclosed that in 30 days, the initiative identified more than 10,000 high- or critical-severity vulnerabilities in software classified as systemically important — operating systems, critical infrastructure control software, and core open-source libraries. The Glasswing expansion on June 2 extended coverage to additional critical infrastructure sectors including energy and financial services. Anthropic also connected Claude to 28 security platforms through a new Claude Compliance API in late May 2026. The intelligence agency story adds an ironic layer: despite Anthropic being placed on a national security concern list (due to Mythos’s offensive capability potential), US spy agencies have not stopped using Claude in active operations.
Why Engineers Should Care
10,000 high- or critical-severity vulnerabilities in 30 days is a discovery rate that no existing SAST, DAST, or human penetration testing program achieves at comparable scale or cost. The bottleneck has shifted from discovery to remediation: finding vulnerabilities faster than development teams can patch them creates its own risk if triage and remediation pipelines aren’t scaled proportionally. For security engineers: the Compliance API connecting Claude to 28 security platforms means you can now integrate Mythos-class analysis into existing SIEM, ticketing, and vulnerability management workflows without building custom integrations. The intelligence agency usage paradox is a procurement signal — even agencies with the deepest security concerns about Anthropic’s capabilities have concluded Claude is operationally necessary.
The Technical Detail Worth Knowing
Anthropic has not disclosed the specific vulnerability classes found, but Hacker News coverage of the Glasswing report noted that the 10,000+ figure is not deduplicated across partners. Each partner organization independently scans its own codebase; the aggregate is a count across all participating organizations. The Claude Compliance API uses a constrained model configuration that restricts output to structured vulnerability reports (CVSS scores, affected component, reproduction steps, remediation guidance) and blocks free-form code generation. This architectural constraint is important: it means the API can be deployed in regulated environments (FedRAMP, SOC 2 Type II) that would otherwise prohibit general-purpose LLM access. The 28 integrated security platforms include at least Splunk, CrowdStrike Falcon, Jira, and ServiceNow — per partner documentation published alongside the Glasswing update.
What Comes Next
Anthropic plans to publish a detailed technical paper on Glasswing methodology and false-positive rates in Q3 2026. The June 2 expansion to energy and financial infrastructure sectors means regulatory disclosure requirements (NERC CIP, SOX, DORA) will start intersecting with the program. Expect the remediation bottleneck to generate demand for AI-assisted patch generation — the logical next phase of Glasswing.
BOTTOM LINE
Claude Mythos just ran the fastest large-scale vulnerability audit in documented security history. The discovery rate is not the constraint anymore — your patch pipeline is.
KEY TAKEAWAYS
- 10,000+ critical vulns in 30 days means the remediation pipeline, not discovery, is now the bottleneck — scale your triage and patch workflows accordingly.
- The Claude Compliance API’s constrained output mode (CVSS + reproduction + remediation only) is the pattern to adopt for deploying AI in regulated security environments.
- Intelligence agencies continuing to use Claude despite security designations signals that operational necessity is outweighing governance concerns — a pattern that will repeat across regulated industries.