Anthropic Says Project Glasswing AI Has Flagged 10,000 High-Severity Bugs in a Month

Key points

• Anthropic says Project Glasswing has surfaced more than 10,000 high- or critical-severity vulnerabilities since launching in October.
• The company has not named the affected projects and has not broken down how many findings are confirmed, patched, or still open.
• AI-generated bug reports have overwhelmed open-source maintainers before; Anthropic has not disclosed Glasswing's false-positive rate.
• Coordinated disclosure through standard CVE channels is underway; no CVEs tied to the program have published yet.
• Glasswing's count is the largest published claim in AI-assisted vulnerability research to date.

San Francisco, Anthropic disclosed Friday that Project Glasswing, an internal program that runs its Claude models against critical open-source software, has found more than 10,000 high- or critical-severity vulnerabilities since launching last month. The company described it as a defensive effort to harden code it considers "systemically" important to the global software supply chain.

What did Glasswing actually find?

Ananthropic has not said. It declined to name the affected projects, citing coordinated disclosure, and won't break down how many of the 10,000 findings have been confirmed by maintainers, patched, or left open. That gap is the one that matters. AI-generated vulnerability reports have flooded open-source maintainers over the past year, with projects including curl and Python publicly complaining about low-quality submissions eating triage time. "Glasswing is already identifying serious issues at a scale that would be impossible for human researchers alone," Anthropic said in its announcement, without disclosing a false-positive rate.

When the first embargoes lift, maintainer responses will tell us whether the count holds up or collapses.

Should you worry about how this compares to other AI bug hunters?

Glasswing isn't alone. Google's Big Sleep, a joint DeepMind and Project Zero effort, reported its first real-world memory-safety find in SQLite last year. DARPA's AI Cyber Challenge closed its final round at DEF CON in August, awarding $4 million to Team Atlanta. XBOW briefly topped the HackerOne U.S. Leaderboard in June. None of those programs have published a raw finding count close to 10,000. Whether that reflects Glasswing's genuine reach or looser counting is the question Anthropic hasn't answered.

We've been tracking this program since [our first report on 28 May](/ story/ai-uncovers-23000-vulnerabilities-in-open-source-software), which noted that Anthropic had not yet said publicly whether any confirmed findings were exploited before disclosure. That question is still open.

How are the findings being disclosed?

Anthropoc said findings are being routed to affected projects through standard CVE channels before public disclosure. No timeline was given. It also didn't say whether Glasswing scans a fixed repository list or selects targets dynamically by dependency exposure, a distinction that affects how evenly risk is covered across the supply chain.

The disclosure lands as Anthropic competes with OpenAI and Google for enterprise security contracts, a market where raw vulnerability output is becoming a visible sales metric. The Claude enterprise security integrations we covered on 26 May point in the same direction: Anthropic is building a story where Claude finds problems and then helps fix them inside the same stack.

What to watch

The number is large enough to matter if it's real and the quality holds. What I'd watch is the first public CVE batch, specifically the severity distribution and how many carry maintainer acknowledgements. A 10,000-finding program with a 30 percent confirmation rate is still meaningful. One with single-digit confirmation tells a different story entirely.