Project Glasswing grows as AI for cyber defense has found more vulnerabilities than the industry expected
AI in security is often presented with a double promise: it can defend better, but it can also magnify attacks. Anthropic's announcement on June 2, 2026 is interesting because it assumes that this ambiguity has already left the theoretical terrain. The company has expanded Project Glasswing, an initiative aimed at protecting critical software with controlled access to the Claude Mythos Preview, from approximately 50 initial partners to approximately 150 organizations in more than 15 countries. ## What happened According to Anthropic, Glasswing's first partners have already found more than 10,000 high or critical severity flaws in their codebases. Based on these first weeks of use and conversations with the security sector, open source maintainers and the United States government, the company decided to expand the group. The new wave includes sectors that were initially underrepresented, such as energy, water, health, communications and hardware. The ad makes two strong claims. The first is factual: models of the Mythos class are already being used at scale to scan critical software and also to write patches, perform pre-release checks and accelerate defensive workflows. The second is a projection: in 6 to 12 months, other companies may have models with similar cyber capabilities and perhaps release them without equivalent safeguards. This second part is Anthropic's inference, not a consolidated fact, but it helps to understand the urgency of the movement. ## The technique behind Cyber ​​defense has a classic bottleneck: discovering vulnerabilities has become just the first step. The hard work is validating, prioritizing, reporting, fixing and deploying fixes without breaking production. Anthropic explicitly admits this by saying that the new bottleneck is already review, dissemination and patching. This changes the role of AI. Instead of just acting as a detector, it needs to participate in the entire cycle. Models like Mythos Preview are useful precisely because they can reason about broad code, identify plausible exploits, and suggest more targeted fixes. The company also connected Glasswing to the launch of Claude Security, a commercial product based on newer public models, and the sharing of tools used by trusted partners. In other words: the closed laboratory starts to become an operational stack. From a scientific point of view, advancement here depends on two combined capabilities: semantic analysis of large code bases and the ability to maintain coherence in long chains of inference. Traditional security tools already meet many standards. The difference between frontier models lies in correlating context, understanding less explicit surfaces and proposing exploration or correction with a more systemic notion. ## Why this matters Glasswing is relevant because it suggests that defensive AI is crossing an operational threshold. We're not just talking about copilots for writing regex or summarizing CVEs. We are talking about models used by organizations whose failure could affect more than 100 million people, according to Anthropic itself. If this is sustained, software security could enter an era of much more intense and continuous screening. At the same time, the announcement reveals a structural shock. Automated fault discovery can grow faster than human ability to fix. This creates an asymmetric race: if defenders don't reorganize processes and infrastructure, simply increasing visibility into vulnerabilities can lead to queues, not security. AI only really helps when it also accelerates disclosure, prioritization and correction. ## The future it anticipates The plausible future is more agentic security, with models running continuous scanning, bug repro, patch proposal, and pre-release verification, always with humans in the approval loop. It is also plausible that open source maintainers and critical infrastructure operators will come to rely on tiered access programs for more powerful models, at least until safeguards mature. There is also a geopolitical implication. When Anthropic prioritizes essential organizations and talks about international expansion, it is treating cyber capacity as a problem of digital sovereignty and infrastructure, not just a product. This raises the debate over who gets access, with what controls and at what pace. ## What to watch out for The first point to monitor is the conversion rate from discovery to deployed fix. Finding 10,000 faults is impressive; Closing the cycle with quality is the real test. The second is governance: what criteria define “trusted organization” to access more sensitive capabilities? The third is diffusion. If models with similar power become widely available without sufficient defensive layers, the gain for defenders may be temporary. Even so, Glasswing already signals something important: the discussion about AI and cybersecurity has entered a less abstract phase. The industry is no longer asking whether models can find serious vulnerabilities. It's asking how to organize institutions, processes and accountability when they start to find too many vulnerabilities.
Sources
- https://www.anthropic.com/news/expanding-project-glasswing
- https://www.anthropic.com/product/security