Anthropic Expands Project Glasswing to Secure Global Infrastructure with AI

Anthropic is expanding its Project Glasswing initiative to 150 new partners to secure critical global infrastructure using the Claude Mythos Preview model. The program has already identified over 10,000 major vulnerabilities and is now focusing on automating the patching and disclosure process. This move is a strategic effort to prepare the cybersecurity industry for a future where powerful AI models are widely available to both defenders and attackers.

Key Points

• Project Glasswing is expanding to 150 additional partners to protect critical infrastructure and open-source codebases used by millions.
• The Claude Mythos Preview model has already proven effective by uncovering more than 10,000 critical security flaws in its initial deployment.
• Anthropic is shifting focus from mere vulnerability discovery to accelerating the patching, disclosure, and remediation process to address industry bottlenecks.
• The company warns that 'Mythos-class' models with powerful cyber capabilities will likely be released by others within 6-12 months, necessitating rapid defensive adaptation.
• Long-term goals include establishing new industry standards for AI-driven security and developing safeguards that allow for the safe general release of advanced cyber-capable models.

Sentiment

The overall sentiment is skeptical and somewhat negative toward the article's framing. HN does not reject the possibility that Project Glasswing and Mythos are useful, and there are serious defenses of cautious deployment and AI-assisted vulnerability research. But the center of gravity is distrustful of Anthropic's marketing, concerned about false positives and organizational misuse, and focused on the gap between impressive claims and reproducible, maintainable security outcomes.

In Agreement

• AI-assisted vulnerability research can be a real defensive advantage when used by skilled teams, especially for finding subtle issues and chaining them into practical exploits.
• A limited rollout may be prudent because public access to highly capable cyber models could accelerate misuse before critical projects have time to patch.
• Project Glasswing could help surface long-neglected security debt and force leadership to treat software risk as a concrete operational issue.
• The value may come from combining frontier models with project-specific harnesses, workflows, and repeated analysis rather than from a simple chat-style model interaction.
• Several commenters accept that even noisy tools can be worthwhile when expert reviewers can quickly reject false positives and keep the occasional meaningful finding.

Opposed

• Many commenters argue that AI vulnerability scanners generate excessive false positives and irrelevant findings, creating more triage work than useful remediation.
• Skeptics see the gated rollout as a marketing, scarcity, enterprise-sales, or capacity-management strategy dressed up as safety.
• Critics object that Anthropic's claims are hard to verify while Mythos access is restricted, making benchmarks, testimonials, and AI-derived patches difficult to trust.
• Some argue that the key bottleneck is exploitability and maintainable fixes, not producing more theoretical vulnerability reports.
• Others worry about handing critical infrastructure and source code to an American AI vendor with government ties, especially for non-US organizations.
• Open-source maintainers push back on AI-generated rewrites or huge patches, warning that they can be disrespectful, unmaintainable, and potentially unsafe without deep domain ownership.