TL;DR
Anthropic is expanding Project Glasswing from an initial group of about 50 partners to about 150 new organizations. The move follows reports that early partners found more than 10,000 high- or critical-severity flaws, shifting pressure from detection to verification, disclosure, patching and deployment.
Anthropic is expanding Project Glasswing, its collaborative AI cybersecurity program, to about 150 new organizations after an initial group of roughly 50 partners found more than 10,000 high- or critical-severity flaws in software codebases, according to source material describing the program.
The expansion moves Project Glasswing beyond its first cohort, which received access in early April to Claude Mythos Preview and began scanning codebases for vulnerabilities. The new group spans more than 15 countries and adds representation from power, water, healthcare, communications, hardware and vendor ecosystems, according to the source material.
The reported focus is no longer only finding flaws. The source material says Anthropic is shifting the program toward verifying findings, coordinating disclosure, writing patches, running pre-release checks, penetration testing and rebuilding some legacy code in memory-safe languages.
Participation is described as gated: new organizations must meet Anthropic security requirements before receiving access. The source material says many partners serve critical infrastructure or provide code relied on by governments and other organizations worldwide.
The bottleneck moved — from finding flaws to fixing them
50 partners found 10,000+ critical vulnerabilities in weeks. So the constraint is no longer detection — it’s verify, disclose, patch, deploy. Anthropic is expanding Project Glasswing to ~150 organizations, and pivoting its weight toward the new chokepoint.
From 50 partners to ~150 — aimed at the leverage points
Not just more headcount. The new group reaches sectors the first cohort underrepresented, and leans toward vendors whose code sits under thousands of downstream systems.
each must meet Anthropic’s security requirements first
cybersecurity vulnerability scanning tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Cute-Patch It Works on My Machine Meme Embroidered Iron on sew on Patch Funny Emblem Programmer Humor
Size: 3 inches tall
As an affiliate, we earn on qualifying purchases.
Finding used to be the hard part
For the whole history of the field, detection was the scarce, skilled work — the chokepoint. A model that surfaces 10,000 critical flaws in weeks inverts that. Toggle before/after and watch the bottleneck move.
The defensive pipeline — where the constraint sits
Same five stages. The chokepoint slides downstream.
Cute-Patch It Works on My Machine Meme Embroidered Iron on sew on Patch Funny Emblem Programmer Humor
Size: 3 inches tall
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
CompTIA CySA+ Certification Kit: Exam CS0-003 2025-2026: A Complete Cybersecurity Analyst Study System for Mastering Threat Detection, Incident Response and Security Monitoring With 1000 practice
As an affiliate, we earn on qualifying purchases.
AI redeployed downstream — and pushed beyond the cohort
Glasswing is consciously shifting its weight from finding toward disclosing, fixing & deploying. The same model helps at the new bottleneck.
Defensive tasks Mythos-class models now take on
Beyond scanning — the work that actually closes the gap.
Writing patches
Partners use the model to fix what it finds — not just flag it.
Pre-release checks
Preventing vulnerabilities from appearing in the first place.
Penetration testing
Simulating attacks to see how a flaw might be exploited.
Rebuilding in memory-safe languages
Attacking whole vulnerability classes at the root.
Claude Security
Uses public frontier models like Claude Opus 4.8 to scan codebases & suggest patches.
The Glasswing tooling
The vuln-finding tools, to trusted security teams — so partners’ methods replicate widely.
code vulnerability detection software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why the urgency is named, not gestured at
The program’s tempo is the tempo of a race against diffusion. Anthropic puts a number on the deadline.
Within 6–12 months, many other labs will have Mythos-class models — and could release them without safeguards.
In that world, cyberattacks could occur much more often, and in much more unpredictable forms. The strategic theory of the whole program: build the defensive head start now, while the capability is still scarce and gated — so when it’s cheap and everywhere, defenders already stand on higher ground.
Capability is scarce & gated
Mythos-class power sits with vetted Glasswing partners under Anthropic’s requirements.
Capability goes ambient
Other labs ship Mythos-class models — possibly ungoverned. The window to prepare closes.
CompTIA CySA+ Certification Kit: Exam CS0-003 2025-2026: A Complete Cybersecurity Analyst Study System for Mastering Threat Detection, Incident Response and Security Monitoring With 1000 practice
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Read it with its difficulties in view
Several are real — some Anthropic states outright, some inherent to the situation. None cancels the core, but all deserve to be held.
Dual use — and the safeguards don’t exist yet
The same capability that finds-and-patches can find-and-exploit. Anthropic says general release needs safeguards that it, and to its knowledge all other developers, have yet to develop. The caution is the clearest evidence of the power.
Gated, even as the logic demands breadth
Advanced defensive capability is allocated by one company’s selection — yet the announcement’s own case is that hundreds of thousands will need access. “Must be gated for safety” sits in tension with “must be widespread to work.”
Not a neutral observer
A frontier lab is at once warning of the danger, helping constitute it, and selling the response (Claude Security, the tooling, the Cyber Verification Program). The warning isn’t wrong — but the commercial frame is worth holding alongside the public-interest one.
Toward a permanent advantage for defenders
Cybersecurity has long been asymmetric in the attacker’s favor — defenders close every hole, attackers need one. The north star is to flip that.
More essential infrastructure
Plus critical-OSS maintainers & safety testers, US & overseas.
Cyber Verification Program
Mythos-class capability for specific cyberdefense tasks — breadth without waiting on full-release safeguards.
Make all software secure
And help the industry adjust how AI changes the core assumptions of cybersecurity.
Reading it in proportion
- The core is hard to argue with: AI made finding cheap & abundant; the bottleneck genuinely moved to patching & deployment; redirecting effort there is sane.
- The caveats sit alongside, not against: one company’s program, one company’s gate, a timeline & products that company has reason to advance — and admittedly-missing release safeguards.
- Hold both halves: the danger is plausible and the 10,000 flaws are real; the response is reasonable and commercially convenient; the aspiration is worthy and unproven.
Why It Matters
The expansion matters because the volume of AI-assisted vulnerability discovery can create a new pressure point for defenders. If thousands of serious flaws can be found in weeks, the harder task may become confirming which findings are real, notifying affected maintainers, fixing the defects and deploying updates before attackers can exploit them.
The program also points to a wider question for the software industry: whether AI security tools can reduce risk faster than they increase the number of discovered, unresolved flaws. Anthropic’s stated approach, as reflected in the source material, is to build defensive capacity while the most capable models remain limited to vetted partners.
Background
Project Glasswing is described as Anthropic’s effort to help secure high-value software systems through collaboration with external partners. The first phase gave about 50 organizations access to Claude Mythos Preview for vulnerability scanning.
The new phase broadens the partner base and shifts attention to downstream remediation. The source material frames this as a change in the main constraint: detection has become faster, while patch validation, disclosure and deployment remain labor-intensive.
What Remains Unclear
Several details remain unclear from the source material, including the names of most participating organizations, how many of the reported flaws have been independently verified, how many have been patched, and whether any were already known before the scans. It is also unclear how Anthropic will measure whether Project Glasswing reduces real-world exploitation risk.
What’s Next
The next phase is expected to focus on onboarding the new organizations, expanding trusted access to Glasswing tooling, and improving disclosure and patching workflows, including for open-source maintainers. The main milestone will be whether the program can move verified fixes into production faster than new AI-assisted findings accumulate.
Key Questions
What is Project Glasswing?
Project Glasswing is Anthropic’s collaborative cybersecurity program aimed at finding and fixing vulnerabilities in important software systems using Claude-based security tools.
What changed in the expansion?
The program is moving from about 50 initial partners to about 150 new organizations, with added focus on infrastructure sectors and vendors whose software affects many downstream users.
Are the 10,000 flaws confirmed as exploited?
No. The source material says partners found more than 10,000 high- or critical-severity flaws, but it does not say those flaws were exploited or fully patched.
Why is Anthropic focusing on patching now?
The reported scale of findings creates a remediation backlog. The program is now aimed at verification, disclosure, patch writing, pre-release checks and deployment.
What remains unknown?
Unknowns include partner identities, patch completion rates, false-positive rates, disclosure outcomes and how broadly similar AI security models will be released by other labs.
Source: Thorsten Meyer AI
