Glasswing: AI-found bugs hit your dependencies

What happened

Anthropic launched Project Glasswing on 7 April 2026, alongside Claude Mythos Preview. AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks and more than 40 other organisations use the model to find vulnerabilities in critical software (Anthropic).

The reported numbers, from Anthropic’s October 2026 update as summarised by Software Improvement Group (SIG):

Partners reported at least 129,000 verified vulnerabilities between April and July 2026, plus 5,500 from Anthropic’s own open-source scanning through October. More than 33,000 were rated critical or high.
External firms reviewed 6,123 candidate findings and confirmed 92.7% of them.
As of 2 October 2026, 6,157 vulnerabilities had been disclosed to maintainers across 591 open-source projects, and 516 (about 8%) were patched.
Anthropic puts the compute cost per vulnerability at under $50.

Maintainers asked Anthropic to slow the disclosure pace because they did not have the capacity to fix bugs at that rate (Cloud Security Alliance).

The counter-view

Raw finding counts overstate the immediate risk. VulnCheck’s audit of the first-month ledger found 23,019 logged findings, 1,611 committed entries and 126 published CVEs at that point, with one confirmed exploited. Across Anthropic-credited and Berkeley datasets, 14 of 1,061 AI-discovered vulnerabilities (1.3%) were confirmed exploited (VulnCheck; VulnCheck 1H-2026).

Both views point to the same practical conclusion: plenty of real bugs, a small and unpredictable share exploited, and a long queue of fixes still to come.

Am I affected?

If you ship software with open-source dependencies, yes, at least indirectly. The questions that matter are how far behind you are and whether you would notice a security release.

bash

# How far behind are your direct dependencies?
npm outdated
pip list --outdated
go list -u -m all
cargo outdated            # needs: cargo install cargo-outdated

# Known advisories across lockfiles, from OSV.dev
osv-scanner scan source -r .

What to do now

If you consume open source:

Shorten the update cycle. Let Renovate or Dependabot open patch-level updates automatically, grouped per dependency family, and merge them when tests pass. A security fix that waits for a quarterly upgrade waits three months.
Track releases, not only CVEs. Many AI-found issues are fixed in ordinary commits, sometimes before a CVE exists or without one. Watch the release notes of your critical dependencies, especially parsers, crypto, compression and anything that handles untrusted input.
Know what is reachable. A bug in a dependency your code never calls, or that never sees untrusted input, can wait. The same bug on an internet-facing path cannot. Rank by exposure, not by count.

If you maintain open source:

Turn on private vulnerability reporting on GitHub, so reports reach you privately rather than in public issues.
Publish a SECURITY.md that says what a valid report must include (affected version, reproduction steps, proof of concept) and how you handle AI-generated reports.
Ask for help early. If a disclosure batch is more than you can fix, say so and ask the reporter for a longer embargo or for help with patches. That is what maintainers did with Glasswing.

Why this is part of the Vulnpocalypse

Glasswing is where the term comes from. It is also the clearest measurement of the Vulnpocalypse’s central gap: AI turned discovery into a cheap, fast pipeline, while fixing still runs at volunteer speed. An 8% patch rate after six months is not a failure by maintainers. It shows the real bottleneck, and it means your dependencies will keep shipping security fixes long after the headlines move on.

Project Glasswing: AI-found bugs are reaching your dependencies faster than maintainers can patch