Anthropic says it found 5,500 verified vulnerabilities in April-October and Glasswing partners found 129K+ in April-July; 33K+ were critical or high severity
Anthropic is expanding a program that allows vetted cybersecurity professionals to test its most powerful AI models with fewer safeguards …
Context & Ripple Effects
Project Glasswing began in April as Anthropic’s effort to apply Claude Mythos Preview to finding and fixing software flaws, then reported more than 10,000 high- or critical-severity findings in May. Its initial partner roster spanned cloud, software and cybersecurity suppliers, including AWS, Apple, Cisco, Google, Microsoft and Palo Alto Networks.
The reported tally accompanies an expanded Cyber Verification Program, which gives vetted security professionals tiered access to Anthropic’s latest models. The scale of verified findings makes the program less a model demonstration than an evidence-generating security workflow.
First-order effects
- Vetted cybersecurity professionals receive broader, tiered access to Anthropic’s powerful models, allowing the company to collect verified vulnerability reports under a more formal program.
- Glasswing’s partner network has a substantially larger pool of reported flaws to validate and route toward remediation, with more than 33,000 classified as critical or high severity.
Second-order effects
- Anthropic’s launch partners face pressure to demonstrate that AI-assisted discovery translates into faster validation and patching rather than an accumulating queue of severe findings.
- The program raises the value of controlled access, verification and responsible-disclosure processes for model providers deploying code-capable systems to security researchers.
Third-order effects
- If comparable programs publish verified outcomes, frontier-model security evaluation is likely to move from aggregate capability claims toward auditable findings and remediation evidence.
- The same model access that improves defensive code review also concentrates dual-use capability in a small set of labs, making access controls and researcher vetting a durable part of the market structure.
The trend: AI cybersecurity is shifting toward controlled, evidence-bearing deployment of frontier models, with verified vulnerability findings serving as the practical benchmark.