Project Glasswing: An Initial Update
4 hours ago
- #Cybersecurity
- #Software Vulnerabilities
- #AI Models
- Project Glasswing launched to secure critical software against threats from advanced AI models.
- Claude Mythos Preview found over 10,000 high/critical vulnerabilities, increasing bug-finding rates by more than tenfold.
- External evaluations show Mythos Preview excels in cyber ranges and exploit benchmarks, outperforming previous models.
- Over 1,000 open-source projects scanned, revealing 6,202 high/critical vulnerabilities, with a high true-positive rate.
- The bottleneck is now verifying, disclosing, and patching vulnerabilities rather than finding them.
- Partners and developers are accelerating patch releases and improving security practices.
- Anthropic released tools like Claude Security and Cyber Verification Program to aid defenders.
- Project Glasswing aims to expand to more partners and eventually release Mythos-class models with stronger safeguards.