世界のソフトウェアを守るProject Glasswing始動 An initiative to secure the world's software | Project Glasswing

Anthropic has unveiled Project Glasswing, an initiative aimed at using AI to discover and remediate vulnerabilities in the world's software at scale. As advances in AI cut both ways for attackers and defenders, the company frames the project as part of a broader effort to tilt the balance decisively toward defenders.

While Anthropic has not yet disclosed full operational details, Project Glasswing appears designed to harness the reasoning capabilities of large language models such as Claude to accelerate code analysis, vulnerability detection, and the generation of automated patches. Traditional security auditing relies heavily on manual review by specialists, an approach that struggles to keep pace with the size and churn of modern codebases — particularly across the sprawling open-source ecosystem on which most commercial software depends. If AI agents can perform continuous, large-scale auditing and propose viable fixes, the gap between vulnerability disclosure and patch deployment could narrow significantly, potentially shrinking the window in which zero-days can be exploited.

The initiative arrives amid a wave of similar research efforts. Google's Project Naptime and its successor Big Sleep have already demonstrated that LLM-driven agents can uncover previously unknown bugs, including a memory safety flaw in SQLite that the Big Sleep team disclosed last year. OpenAI has published work on automated vulnerability discovery, and DARPA's AI Cyber Challenge has spurred teams to build end-to-end systems that find and patch flaws in real-world code. Collectively, these efforts suggest that LLMs are becoming a meaningful complement to established techniques such as fuzzing, symbolic execution, and static analysis, rather than a replacement for them.

The defensive urgency is heightened by the parallel rise of offensive AI. Reports from threat intelligence groups, including Anthropic's own misuse analyses, indicate that adversaries are using LLMs to refine phishing campaigns, accelerate reconnaissance, and assist with malware development. In that context, automating the defensive side of the equation is no longer optional. Proponents argue that defenders enjoy a structural advantage with AI: they have legitimate access to source code, build systems, and telemetry that attackers typically lack, and that asymmetry could be amplified by capable agents.

Project Glasswing also fits within Anthropic's Responsible Scaling Policy, under which the company evaluates the cyber capabilities of its frontier models and considers the dual-use implications of releasing them. By channeling those capabilities into a defensive program, Anthropic appears to be operationalizing a thesis it has articulated repeatedly: that frontier models should be deployed in ways that produce concrete safety dividends, not just abstract assurances.

Key questions remain open. It is not yet clear how Glasswing will be delivered — whether as a hosted service, a set of tools for security teams, partnerships with critical-infrastructure operators and open-source maintainers, or some combination of these. The scope of code coverage, disclosure practices, and how findings will be coordinated with upstream maintainers will all be important indicators of the project's real-world impact. There are also unresolved challenges around false positives, patch quality, and the risk that vulnerability-finding agents could be repurposed by adversaries if their methods leak.

Broader industry coordination is likely to determine whether efforts like Glasswing translate into measurable reductions in exploitable vulnerabilities. The open-source community, cloud providers, and government agencies such as CISA have all pushed for more proactive, automated approaches to software assurance, and AI-driven auditing is increasingly seen as a natural extension of that agenda. If Anthropic can demonstrate consistent, high-signal results — and share them in ways that strengthen the wider ecosystem rather than just its own products — Project Glasswing may become a notable test case for the proposition that frontier AI can make software meaningfully safer.