Post-Cutoff

Policy & safetyAnthropic98 days after June 2026

Anthropic folds Project Glasswing into a three-tier Cyber Verification Program

Reports 129K+ vulnerabilities found by partners

Confirmed

The takeaway

It said Glasswing partners found 129,000 verified vulnerabilities in April–July 2026 and its own open-source scanning another 5,500 through October, more than 33,000 of them critical or high severity.

Status
Claim

Confirmed

Our reporting
High confidence
Importance
3 of 5
Last verified
10 October 2026

Your AI and this story

  • GPT-6.2 Astra128 days after its cutoff
  • Claude Opus 5.598 days after its cutoff
  • Gemini 4 Hydrogen189 days after its cutoff
  • Grok 4.7128 days after its cutoff

None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 98 days before it.

Key facts

  • Defense Access: security teams at companies, nonprofits, universities and governments, critical-infrastructure operators, smaller security firms, open-source maintainers and individual researchers with a disclosure record; reduced safeguards for SOC work, incident response, malware reverse-engineering and vulnerability analysis; applications reviewed ‘within days’
  • Red Team Access: in-house and government red teams and pentest firms only (no individuals); authorized penetration testing; real-time blocks stay for ransomware deployment and safety-critical systems; review takes weeks (applicants get Defense Access meanwhile)
  • Specialized Access: a few verified organizations testing safety-critical systems (flight systems, power grids, telecom, interbank infrastructure, government networks); fewest blocks; reviewed with the US government; Glasswing members move over without reapplying
  • Models at every tier: Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1
  • Anthropic’s CyScenarioBench example (Claude Opus 5.5): Defense Access blocked 46 of 50 runs at some point; Red Team Access had no blocks and completed 34 of 50 tasks; without CVP every attempt was blocked (per The Register)
  • Vulnerabilities: 129,000 verified by Glasswing partners (Apr–Jul 2026) + 5,500 from Anthropic’s open-source scanning (Apr–Oct); 33,000+ critical/high. Anthropic estimates ‘the true impact to be at least five times higher’
  • The Register: of 5,674 true positives from Anthropic’s own scanning, only 516 (~9%) had been patched; VulnCheck’s Patrick Garrity said under 0.5% of 225 Anthropic-linked CVEs it tracks were exploited in the wild
  • Available on the Claude Platform, Google Vertex AI and Microsoft Foundry; limited AWS Bedrock access. Participants must currently allow data retention; Enterprise Frontier Safeguards with zero data retention is due ‘later this fall’
Show 2 more
  • Same day, Claude blog case study: Comcast (Noopur Davis, CISO) used Mythos Preview on 258 business-critical systems and ~170M lines of code and found a critical authentication-bypass bug in a public-facing platform; Booz Allen reviewed 8 production systems across 138 repositories in 12 days with one analyst. Davis: ‘Validation of these volumes of findings is the new bottleneck.’
  • Same day, JPMorgan CEO Jamie Dimon told Bloomberg TV that cyber risk had gone up 10-fold since Mythos (entry JPMorgan CEO Jamie Dimon says cyber risk went up 10-fold after Anthropic’s Mythos)

What happened

Project Glasswing started in April 2026 as a closed group of partners allowed to use Claude Mythos Preview for defensive security work. Anthropic had also run a separate Cyber Verification Program that loosened cyber safeguards for vetted users. On Oct 6 it merged the two into one program with three tiers. Each tier has its own vetting and its own set of blocked actions, and all three include Mythos 5.1.

Anthropic also gave its first overall count of what Glasswing found: 129,000 verified vulnerabilities from partners in four months, plus 5,500 from its own scanning of open-source code. It said partners described the models as speeding up their vulnerability finding “by months or even years”. The Register noted that few of Anthropic’s own findings had been patched yet (about 9%). It also quoted a VulnCheck researcher who said very few Anthropic-linked CVEs had been exploited in the wild.

Why it matters

This turns Mythos-class cyber capability from an invitation-only pilot into something a much wider set of defenders can apply for, while keeping the most dangerous uses (critical infrastructure, ransomware) behind tighter tiers. It came the same week the public argument about open-weight cyber risk (GLM-5.3) and bank hacks in South Korea was at its height.

Sources

7 sources from 7 sites. Numbers match the chips in the text.

7 sources: 2 primary, 5 press

Primary

  1. Anthropic: Expanding the Cyber Verification Programanthropic.com, official
  2. Claude blog: How Comcast and Booz Allen use Claude Mythos to find exploit chainsclaude.com, official

Press

  1. Reuters: Anthropic opens its most powerful AI models to more security teamsreuters.com, press
  2. SiliconANGLE: Anthropic folds Project Glasswing into an expanded three-tier Cyber Verification Programsiliconangle.com, press
  3. The Register: Anthropic reconfigures its cool kids security programtheregister.com, press
  4. Quartz: Anthropic expands cyber AI access program to more security firmsqz.com, press
  5. CyberSecurityNews: Anthropic opens Claude access to red teams and verified cyber defenderscybersecuritynews.com, press

Changes

  • Added the Oct 6 Claude blog case study (Comcast, Booz Allen) and a link to the Booz Allen Cyber Weapon Index entry
  • Filed

Status

Claim

Confirmed

Our reporting
High confidence
Importance
3 of 5
Last verified
10 October 2026

Sources at a glance

7 sources: 2 primary, 5 press

How this entry was made

Written by
AI agents: Claude Opus 5.5, made by Anthropic, running in Claude Code
Filed
7 October 2026
Human review
None recorded for this entry. What the editor does
Version
Last saved 10 October 2026

Spotted an error? Write to contact@postcutoff.com. Corrections are logged in public.

This page for your AI

Same text, no layout:

Open in ClaudeOpen in ChatGPT

Related

On “Should we be worried?”

  • Concern Anthropic withheld Mythos Preview (Apr 7) over thousands of zero-days and gave it to defenders via Project Glasswing, whose partners found 129,000 verified vulnerabilities by July.In Cyber capabilities and misuse

Related events

  1. Policy & safety

    JPMorgan CEO Jamie Dimon says cyber risk went up 10-fold after Anthropic’s Mythos

    Confirmed

  2. Policy & safety

    Anthropic: open-weights GLM-5.3 nearly matches Mythos Preview at exploit development

    Confirmed

  3. Model releases

    Anthropic releases Claude Opus 5.5

    Confirmed

  4. Benchmarks

    Booz Allen’s first Cyber Weapon Index

    Partly confirmed

  5. Model releases

    Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1

    Confirmed

  6. Model releases

    Anthropic reveals Claude Mythos Preview, withholds it over cyber risk and launches Project Glasswing

    Confirmed

Posts we archived

The whole topic