Anthropic expands Glasswing for AI vulnerability defence

Anthropic has expanded Project Glasswing by inviting around 150 more organisations in more than 15 countries to test Claude Mythos Preview. The programme now covers power, water, healthcare, communications, hardware and software maintainers, while extending from vulnerability discovery into remediation, trusted tooling and open-source disclosure. The key signal is that AI-assisted security is increasing the pressure on verification and repair, not only detection.

• Initial members found 10,000+ severe flaws in systemically important software
• The project shifts from discovery toward verification, disclosure and repair

The fact

Anthropic has invited around 150 more organisations in more than 15 countries to join Project Glasswing and test Claude Mythos Preview. The new intake spans power, water, healthcare, communications, hardware and software maintainers whose codebases support governments and critical organisations. Anthropic says the first 50 members used Mythos to find more than 10,000 high-or-critical-severity vulnerabilities, while the programme is now extending into remediation, trusted tooling and open-source disclosure.

The Assessment

The expansion shows the cyber security bottleneck moving from vulnerability discovery to verification and repair. Anthropic is not simply widening access to a powerful model; it is trying to wrap Mythos inside a controlled defence workflow before similar capabilities become more widely available. For infrastructure operators and software maintainers, the signal is practical: AI-assisted security will increase patching pressure as much as detection capacity.

What to Watch

Watch whether Glasswing becomes a repeatable remediation channel, and whether other AI labs release Mythos-class cyber models within the next 6–12 months with weaker safeguards.

Also read: Bergen Engines wins Liberty Energy 500MW AI power order

Also read: Vodafone expands connected vehicle partnership with Geely