Only 442 of the nearly 38,000 vulnerabilities claimed to have been found using Anthropic's Project Glasswing have received a CVE identifier, and only two of those have been exploited in the wild.
https://www.vulncheck.com/blog/anthropic-glasswing-november-update
@campuscodi evidently, 30 thousand zerodays
Maybe. We don't know. 30,617 findings that they haven't disclosed.
@piggo @campuscodi
@campuscodi That gap between 38,000 claimed findings and 442 CVEs is wild — and only two exploited in the wild. Really makes the headline number feel a lot less scary once you see the funnel.
@YoungImpAngel @campuscodi the exploited in the wild means antrophic is faster than humans at a rate 442 to 2. The 38k to 442 is the ratio of trash ppl misunderstanding tech to professionals.
@splinux @YoungImpAngel @campuscodi No, it means Only 442 of the nearly 38,000 vulnerabilities claimed to have been found using Anthropic's Project Glasswing have received a CVE identifier, and only two of those have been exploited in the wild.
You are making assumptions about the exploitation conditions (actual users of the affected versions, other conditions necessary for the vulnerability...)
@karlggestd @splinux @YoungImpAngel @campuscodi Serious Qs: Who or what process grants a CVE id? Also is the low in the wild number indicative of many of these vulns being associated with older little used libraries that no longer have high exploitability?
@VexaciousFacts
Of course
@splinux @YoungImpAngel @campuscodi
Nov 12, 2026, 12:02 PM@michael @YoungImpAngel @campuscodi @splinux
There is no reason to think that an effort capable of verifying 38,000 vulnerabilities wouldn’t have been able to identify 2.
It makes no sense to generate massive numbers of error reports that consist, for the most part, of false positives. That’s like thinking it’s reasonable to check the time on a broken analog clock—after all, it’s right twice a day, isn’t it?
Nov 12, 2026, 12:17 PM@michael @YoungImpAngel @campuscodi @splinux
As you rightly point out, there are 37,558 false positives or duplicates.
Regarding the remainder, we know they were accepted as CVEs, but we do not know:
* the risk rating for each of those 442 entries;
* which ones affected active library versions;
* the conditions required to exploit the vulnerability.
1/2
@michael @YoungImpAngel @campuscodi @splinux
As you also rightly note, it is remarkable to assume that—given these tools can identify so many potential exploits—the very people who create such exploits wouldn't be using these same tools themselves.
In the wild: 2
2/2
@michael I believe the exploited number is meaningful because after it was published or patched there is usually enough public information about it to write an exploit. I believe there are documented cases of exploit writers vibe coding exploits for recent published CVEs pretty quickly. (Even a silently released patch reveals a lot of details.)
Exploits for already-patched vulns are still useful because so many organisations are very slow at patching or even never patch. It's not rare to see news stories about new exploits for RCE bugs that have had patches available for months to years, still managing to pwn a few thousand installations.
So if a given vuln is useful to miscreants, you should expect there to be exploitation of it in the wild.
(The only time I would expect timely patching to completely erase the value of a serious vuln is when the vuln is in software that only one organisation runs - like SaaS products where patching is completely under the vendor's control.)
I 🩷 SANKEY flow charts
Wondering how much fossil fuel was burned to find those vulnerabilities.
The ROI on two cve's seems a little low. /s
What does it mean to be not in the ledger?
That makes it sound like Glasswing discovered ~30k novel vulnerabilities, which would be something to brag about.
@campuscodi Thanks, I needed this.
Finding security bugs with #AI is like throwing spaghetti at the wall and see what sticks, something stuck but it's only a bit.
@hnapel @campuscodi The problem for all is that the low percentage doesn't really matter, a small percentage of a really large number is still large and the real findings from all the AI research is orders of magnitude more than previously.
Heck, even the number of issues I just opened for fixing in public (opened recently, submitted over the past few months) is more than for whole years previously.
If they are not getting validated then perhaps their competitors have better models!
@hnapel @campuscodi But mostly, as the article states, the bottleneck remains the human triage and validation, which is both expensive and outsourced to the projects.
@campuscodi The hype gets more damaging as the billionaires leading these companies get more powerful.
They also talk like politicians nowadays, which says a lot.
Nov 12, 2026, 2:56 PM@campuscodi 38'000 vulnerabilities, 277 (sic) fixed. :thisisfine:
Somehow looking at the list of vulns, is not much better than just looking at the source code, where all the vulns are anyway.
@campuscodi I think the word "vulnerabilities" belongs in scare quotes when coming from Anthropic!
Nov 12, 2026, 5:03 PM@campuscodi Great chart, really clear!
@campuscodi
#altText
A graph displayed in nightmode titled "anthropic glasswing findings"
A flow chart showing all reported vulnerabilities splitting into designations as we go left to right. It splits about 20:80 to give "in ledger" and "not in ledger". "In Ledger" (the 20%) splits further into pieces that have widths in proportion to their ratio of the set. Only 193 of the 37,966 reported vulnerabilities were valid and unknown, looking like a hair shaved from a block
A "VulnCheck" logo appears at the bottom right
@campuscodi Also overburdened Open Source projects.
To be clear, AI is finding a lot of real issues, some really serious, many very obscure and lots of denial of service.
But the flood is debilitating. A process that was built for sporadic bursts of high energy and focus can't function under this fire hose.
I do triage on incoming issues for Samba as a personal time contribution and it is exhausting.