Anthropic's AI Finds 10,000+ Critical Bugs — and Now It's Coming for Power Grids and Hospitals
Anthropic has expanded Project Glasswing — its Claude Mythos-powered vulnerability scanning initiative — to approximately 150 new organizations in more than 15 countries, covering sectors such as power, water, healthcare, communications, and hardware that were underrepresented in the first wave. Since launching in April, the program has surfaced more than 10,000 high- or critical-severity security flaws.
When Anthropic launched Project Glasswing in April 2026, the pitch was simple and audacious: give the world’s most powerful AI model access to the code underpinning critical global software infrastructure, and see how many vulnerabilities it could find before attackers did.
The answer, after several weeks of work with an initial cohort of about 50 partner organizations, was: a lot. More than 10,000 high- or critical-severity vulnerabilities were surfaced. Separately, of 1,752 high- or critical-rated findings from an open-source scan that were independently reviewed, over 90% were confirmed as valid. One of the most significant was CVE-2026-5194, a critical WolfSSL flaw (CVSS 9.1) that would have allowed attackers to forge certificates and impersonate trusted services across systems that depend on the widely-used cryptography library.
Now Anthropic is scaling up. On June 2, the company announced it was extending Project Glasswing to approximately 150 new organizations across more than 15 countries — more than tripling the program’s reach — with a deliberate expansion into sectors the initial cohort had not covered.
What Changed in Phase Two
The first wave of Glasswing partners included established technology companies such as Amazon Web Services, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, among others. These are organizations with mature security practices, large internal engineering teams, and the infrastructure to process and triage AI-generated findings at scale.
Phase two deliberately targets organizations that are harder to secure but potentially more catastrophic when breached. New sectors added include power and energy utilities, water treatment and distribution systems, healthcare providers and hospital networks, and telecommunications infrastructure. Many of the new participants are vendors: companies whose software runs inside critical infrastructure systems rather than operating the infrastructure directly — making them high-value attack targets with potentially lower security maturity than the organizations they serve.
Anthropic’s framing for why these organizations were prioritized is explicit in its announcement: for most partners, it estimates that a major attack could affect more than 100 million people.
Sources told CyberScoop that NetSkope and Rubrik, which specialize in cloud security and data management, are part of the group given access in this round.
The Bottleneck Isn’t Finding Bugs. It’s Fixing Them.
One of the more counterintuitive insights to emerge from Phase One of Glasswing is that AI-powered vulnerability discovery has progressed further and faster than anyone expected — but it has also revealed a structural bottleneck that AI cannot yet solve.
Cloudflare found 400 critical and high-severity bugs among 2,000 total findings. Mozilla found 271 vulnerabilities in Firefox 150 alone. An open-source scan across more than 1,000 projects flagged 23,019 potential vulnerabilities, of which 6,202 were estimated as high- or critical-severity; of 1,752 high- or critical-rated findings independently reviewed, over 90% were confirmed as valid.
These are extraordinary numbers. They also create a new problem: the pipeline of confirmed vulnerabilities is generating more remediation work than engineering teams can absorb. As Anthropic noted in its expansion announcement, “The bottleneck in fixing bugs like these is the human capacity to triage, report, and design and deploy patches for them.”
This is what the company describes as the deeper challenge of the program: not the discovery phase, but the response phase. Even with AI doing the detection, patching critical vulnerabilities in widely-deployed, production software requires human judgment, careful testing, and coordination with the open-source maintainers or enterprise vendors who own the affected code. That process cannot be automated away — yet.
Anthropic says that Claude Security, using the publicly available Claude Opus 4.8 model (not the restricted Mythos variant), has already helped patch more than 2,100 vulnerabilities in three weeks. Claude Mythos Preview, by contrast, is kept restricted because Anthropic says it lacks safeguards sufficient to prevent serious misuse.
The Architecture of Controlled Access
Project Glasswing represents a genuine experiment in responsible deployment of a capability-threshold AI model. Claude Mythos was not publicly released when it crossed what Anthropic describes as a critical capability threshold for cybersecurity offense. Instead, Anthropic chose a restricted-access model: partner organizations get API access under specific use-case constraints, and Anthropic maintains oversight of how the model is being used.
The expansion to 150 organizations makes that oversight challenge significantly harder. Fifteen-plus countries means multiple regulatory jurisdictions, varying disclosure norms, and different legal frameworks around responsible vulnerability disclosure. A critical finding in a U.S.-regulated hospital system has different reporting requirements than the same finding in a German energy utility or a South Korean telecommunications provider.
Anthropic says it is managing this through what it calls a “structural advantage for defenders over attackers” framework: the goal is not simply to find vulnerabilities, but to ensure that the time advantage for defenders — their ability to patch before attackers weaponize a finding — is permanently widened. That requires treating disclosure coordination as a core program function, not an afterthought.
A Preview of the AI-Security Relationship
The broader significance of Project Glasswing extends beyond the vulnerability counts. It represents the first large-scale test of a model that has crossed a cybersecurity capability threshold — and a template for how the AI industry might handle future capability milestones.
The implicit argument Anthropic is making is that the right response to building a model powerful enough to cause serious harm in the wrong hands is not to delay deployment or embargo the model entirely, but to deploy it under controlled conditions for defensive purposes, at scale, before adversarial actors can develop equivalent capabilities themselves.
Whether that framework holds at the next capability threshold — whenever models become capable of autonomous, end-to-end exploit development without human guidance — is a question the security community is watching closely. For now, Glasswing offers something that has been largely absent from AI safety debates: empirical evidence of what AI-powered security defense actually looks like in production.
The evidence so far suggests that the defenders are winning, at least in this phase. The race is still very much on.
Correction (November 15, 2026): 8 sentences were corrected or removed to match the sources listed on this page.