Anthropic has announced a new round of expansion for Project Glasswing, its ongoing AI-assisted cybersecurity research program. The move signals continued momentum for an initiative that has grown steadily since its launch, pulling in a wider range of institutional partners and extending the reach of Claude-powered vulnerability research across more sectors.
What the Expansion Covers
Project Glasswing has been one of the more closely watched programs in applied AI security. Built around Anthropic's original Glasswing framework, it uses Claude models to scan software systems for vulnerabilities at a scale and speed that traditional security auditing cannot match. The latest expansion adds new partner organizations to that network, though the full list of incoming participants has not been publicly disclosed.
Key Facts
- Project Glasswing is Anthropic's AI-powered cybersecurity research initiative
- The program uses Claude models to identify software vulnerabilities across partner systems
- Earlier phases surfaced more than 10,000 software flaws within the first 30 days
- Partners have included NATO, Samsung, and critical infrastructure operators
- Anthropic has also granted the EU's cybersecurity agency ENISA access to findings
The scale of what Glasswing has already accomplished gives context to why new partners are lining up. Earlier reporting confirmed that Claude identified more than 10,000 software flaws in a single 30-day period, a figure that caught attention well beyond the cybersecurity community. For organizations managing large codebases or complex infrastructure, the prospect of that kind of automated analysis is hard to ignore.
The expansion of Project Glasswing reflects growing institutional confidence in AI-assisted security research and the practical value it can deliver at scale.Anthropic
A Program That Has Been Building Quietly
Glasswing has expanded in deliberate stages. What started as a contained research effort has since grown into a multi-sector program. A previous expansion brought the partner count to around 200, including organizations like NATO and Samsung, alongside operators of power grid infrastructure. That breadth has made Glasswing one of the more consequential real-world deployments of frontier AI models to date.
Anthropic has also been thoughtful about transparency. The company previously opened up the ability for partners to share findings from the Mythos evaluation framework that underpins the program, a step that allowed the research community broader access to what the system was discovering. The EU's cybersecurity watchdog ENISA was among the bodies granted access, reflecting the geopolitical weight the program has started to carry.
For Anthropic, Glasswing represents a concrete argument that safety-focused AI development and practical utility are not competing goals. Running Claude against real-world software at this scale generates data that feeds back into model development, while also delivering measurable security value to partners. The expansion announced today continues that loop, adding more systems, more findings, and more evidence of what AI-assisted security research can do when applied consistently over time.
Full details on the new partners and the specific scope of this latest phase are expected to follow in subsequent updates from the company. Given the trajectory of earlier expansions, further announcements are likely as Glasswing continues to grow.
