Anthropic Says Project Glasswing AI Has Flagged 10,000 High-Severity Bugs in a Month
The Claude-based scanner has been pointed at widely deployed open-source code since October. Anthropic has not named the affected projects.
San Francisco — Anthropic disclosed Friday that Project Glasswing, an internal program that turns its Claude models loose on critical open-source software, has surfaced more than 10,000 high- or critical-severity vulnerabilities since launching last month.
The company described the effort as a defensive push to harden code it considers "systemically" important to the global software supply chain. Glasswing went live in October. Anthropic said the bug count covers roughly four weeks of operation.
The company declined to name the affected projects, citing coordinated disclosure. It also declined to break down how many of the 10,000 findings have been confirmed by maintainers, how many have been patched, and how many remain open. That distinction matters. AI-generated vulnerability reports have flooded open-source maintainers over the past year, and projects including curl and Python have publicly complained about low-quality submissions consuming triage time.
"Glasswing is already identifying serious issues at a scale that would be impossible for human researchers alone," Anthropic said in its announcement. The company did not disclose the false-positive rate, which is the figure maintainers care about most.
The initiative sits alongside a growing field of AI-assisted bug hunting. Google's Big Sleep project, run by DeepMind and Project Zero, reported its first real-world memory-safety find in SQLite last year. XBOW, a startup running autonomous penetration agents, briefly topped the HackerOne U.S. leaderboard in June. And DARPA's AI Cyber Challenge wrapped its final round at DEF CON in August, awarding $4 million to Team Atlanta for an autonomous patcher built on similar foundations.
Glasswing's scale claim is the largest published to date. Whether it holds up depends on what maintainers say when the embargoes lift.
Anthropic said findings are being routed to affected projects through standard CVE channels before public disclosure. The company did not provide a timeline. It also did not say whether Glasswing operates on a fixed list of repositories or selects targets dynamically based on dependency graphs (a method Google's OSV-Scanner uses to weight by downstream blast radius).
The disclosure follows a August advisory from CISA urging federal agencies to inventory their open-source dependencies. It also lands as Anthropic competes directly with OpenAI and Google for enterprise security contracts, a market where demonstrable vulnerability output is becoming a sales metric.
No CVEs tied to Glasswing have been published yet under the program's name. Anthropic said the first batch of coordinated disclosures is expected in the coming weeks.
