VulnCheck: 7 months of Project Glasswing yields 442 Anthropic-attributed CVEs, 2 exploited in the wild, 217 fixes
- Seven months after Project Glasswing launched, VulnCheck counts 442 CVEs attributable to Anthropic, about 0.5% of the 93,450 CVEs published this year, and only 2 of them (CVE-2026-61500 and CVE-2026-26980) have been exploited in the wild.
- Anthropic's own ledger lists 37,966 findings as of November 2, but only 7,349 (19.4%) are public and just 217 are marked fixed; 1,654 (4.4%) are duplicates and 419 withdrawn.
- Since September Anthropic added nearly 12,000 findings while recording only 15 additional fixes, widening the gap between discovery and remediation.
- On Anthropic's OSS Scanner, curl rejected 19 of 34 published reports as non-vulnerabilities, confirmed 6 as low severity, and left 9 unassessed, which the author says undercuts the claimed 94.2% true positive rate (up from 90.8%).
- VulnCheck tracks more than twice as many Anthropic-attributable CVEs (442) as Anthropic's ledger discloses as fixed (217), drawing on CVE records, GHSAs, Linux kernel commits, and manual advisory audits.
Hacker News 의견들
Honestly the biggest tell is that we haven't had a single catastrophic exploit come out of these findings in a year. LLMs help us patch faster, sure, but I expected at least a couple of supermarket-level bangers by now.
Catastrophic vulnerabilities have been dropping every other day for months, what are you talking about?
There was the PHP thing and the asshole one, and who's to say the supermarket attackers didn't use AI anyway.
The article leans hard on the 19.4% that went public. The other 80% could be unfixed, could be fake, or could still be under embargo, so this is analysis of survivors.
I'd bet the backlog is just human review, not something statistically special about those findings. Anthropic should teach the system to propose the patch too, that's the whole point of having the best coding model.
Or maybe it isn't worth submitting a pile of bullshit to human review. If they were sitting on a gold mine they'd throw people at the queue instead of inflating the numbers.
37,966 findings against 217 fixes. If AI is much better at finding bugs than fixing them, securing anything gets harder, and I hope they didn't spend equal tokens on both sides.
Once there's a documented test for a bug, AI is really good at fixing it, as long as a human reads the patch.
Coding is hard. Any patch it writes probably comes with new bugs, same as human patches do.
Does the author think Microsoft or Cisco are going to credit Anthropic when an internal team uses AI to find a CVE in their own code? Attribution is probably a condition of getting access in the first place.
The PHP2Shell thing was weirdly comforting. Frontier models on PHP produced a bug built on a known logic flaw that's brutal to exploit, and real-world impact stayed small because the chain is long and PHP configs vary.
The conclusion lines up suspiciously well with VulnCheck's business model. Feels like a thinly veiled ad for their prioritization product.
The patch story is what gets me. For PHP2Shell there were fixed .deb packages for the trendy distros and nothing for older ones, and one-line prompts did not flood the world with patched servers.
Expected this to happen. Maintainers buried in issues, most of them nothing burgers.