
VulnCheck's seven-month audit of Anthropic Project Glasswing credits the program with 442 CVEs and places two of them on the VulnCheck Known Exploited Vulnerabilities catalog. Anthropic's disclosure ledger, updated 2 November 2026, lists 37,966 findings and 217 fixes. Of those findings, 19.4 percent are public. The gap between candidate volume and confirmed fixes is what a security lead should weigh before model output enters the backlog.
Key takeaways
- VulnCheck attributes 442 CVEs to Anthropic over about seven months, near 0.5 percent of 93,450 CVEs published in 2026, against 217 fixes on the 2 November ledger.
- Only CVE-2026-61500 and CVE-2026-26980 from that credited set appear on VulnCheck KEV, about 0.45 percent. VulnCheck publishes no product, CVSS, or exploitation write-up for either, and the piece includes no Anthropic reply on those IDs.
- From 8 September to 2 November the ledger went from 26,153 to 37,966 findings and from 202 to 217 fixes. The claimed true-positive rate rose from 90.8 percent in May to 91.4 percent, then to 94.2 percent. This article uses the table duplicate count of 1,670, up from 2. VulnCheck's takeaways say 1,654. Withdrawn findings rose from 245 to 419.
- Of curl's first 34 published reports from a batch of 45, 19 were not vulnerabilities, 6 were confirmed low, and 9 were pending: 55.9 percent false positives on that slice. OSS Scanner reports, launched 8 October 2026, are fully model-generated and may be invalid.
What did Project Glasswing claim on 7 April 2026?
On 7 April 2026 Anthropic launched Project Glasswing and said it had already found thousands of additional high- and critical-severity vulnerabilities and was disclosing them to open-source maintainers. VulnCheck quotes the launch line: “We have identified thousands of additional high- and critical-severity vulnerabilities that we are working on responsibly disclosing to open source maintainers.”
VulnCheck opened a public tracker on 17 April 2026, an open-source repository named Anthropic Credited CVEs. A first post on 15 April covered CVEs attributed to Anthropic researchers and Project Glasswing. Earlier posts are dated 15 April, 9 June, 28 July, and 8 September 2026. The November piece is the fifth in that series, not a one-off reaction to curl.
The tracker uses CVE records, GitHub Security Advisories, Linux kernel commits, Anthropic’s ledger, and manual reads of advisories. The ledger updates in infrequent bulk batches, so VulnCheck kept its own repo between dumps. VulnCheck does not dispute that AI finds vulnerabilities. The author says that story runs past any single frontier model.
A candidate count, a CVE credit, a maintainer confirmation, a fix, and observed exploitation are five different events. The launch folded them into one narrative. Seven months of public outcomes check the sentence about thousands of high- and critical-severity bugs moving through responsible disclosure. Glasswing’s launch framed the work as the Mythos era of vulnerability discovery. VulnCheck’s question is practical: are the credited bugs useful to an attacker, and are they getting fixed.
What do 442 Project Glasswing vulnerabilities actually represent?
VulnCheck attributes 442 CVEs to Anthropic across roughly seven months, about 63 a month, and about 0.5 percent of the 93,450 CVEs published in 2026. A CVE identifier records one publicly disclosed vulnerability. Assignment is not a certification of exploitability, business impact, or patch quality. The 442 figure is a credit count, and a weak stand-in for a risk total.
That credit count is more than double the 217 fixes on Anthropic’s public ledger as of 2 November. VulnCheck counts a vulnerability when Anthropic appears in a CVE record, a GHSA, a Linux kernel commit, or the ledger, then manually checks advisories that automated pulls miss. A fix mark on the ledger is a separate field. The same bug can carry a public credit and still sit unfixed. 442 and 217 are not a before-and-after pair. One counts attribution across public artifacts. The other counts a status flag inside one vendor export.
Two of the 442 appear on VulnCheck KEV: CVE-2026-61500 and CVE-2026-26980, about 0.45 percent of the credited set. VulnCheck publishes no product names, CVSS scores, or exploitation write-up for either CVE. The November piece includes no Anthropic statement answering those two identifiers. Two KEV matches are a lower bound on observed exploitation. Absence from the catalog does not prove a bug was never used.
Inside a company tracker the same split applies. Keep candidate, confirmed, duplicate, fixed, and disclosed in separate columns, and deduplicate on shared root cause before publishing a validation rate. Fund the basics before funding AI findings when the column that grows is candidates, not fixes. Thousands of candidates can exist while public CVE credit stays in the hundreds. A board slide that quotes 37,966 findings as patched criticals is reading a different document than the ledger.
Why the fix counter barely moved
Between the 8 September and 2 November snapshots, findings moved from 26,153 to 37,966 and fixes moved from 202 to 217. That is 11,813 more findings and 15 more fixes. Public entries moved from 2,736 (10.5 percent) to 7,349 (19.4 percent). Withdrawals moved from 245 to 419. The claimed true-positive rate moved from 91.4 percent to 94.2 percent. The May 2026 launch baseline was 23,019 findings and a claimed true-positive rate of 90.8 percent. Updates arrive in infrequent bulk batches, so the curve between snapshots is invisible.
| Ledger field | May 2026 launch | 8 September | 2 November |
|---|---|---|---|
| Findings | 23,019 | 26,153 | 37,966 |
| Made public | not stated | 2,736 (10.5%) | 7,349 (19.4%) |
| Fixed | not stated | 202 | 217 |
| Withdrawn | not stated | 245 | 419 |
| Duplicates | not stated | 2 | 1,670 |
| Claimed true-positive rate | 90.8% | 91.4% | 94.2% |
On 2 November, 217 fixes out of 37,966 findings is about 0.57 percent. That ratio is a snapshot of mixed ages, not a cohort remediation rate. Findings filed in late October have not had the same calendar as findings filed in May. A two-month window cannot show that an unfixed row will never be fixed. It can show the fix counter barely moved while candidates jumped by nearly 12,000. VulnCheck had warned in September that findings left sitting get found and fixed by someone else first. The duplicate jump matches that warning.
VulnCheck’s write-up disagrees with itself on duplicates. The key takeaways say 1,654 duplicates, 4.4 percent of findings. The comparison table says the count moved from 2 on 8 September to 1,670 on 2 November. This article cites the table figure, 1,670, and leaves the 1,654 line as the author’s alternate summary. Both published counts sit well above the 217 fixes.
More findings were withdrawn or marked duplicate than were fixed, yet the claimed true-positive rate rose across the three snapshots. VulnCheck’s author suspects a formula that leaves duplicates and withdrawals out of the denominator. That is a suspicion, not a published formula. The ledger does not show a denominator an outsider can recompute.
VulnCheck reports a display bug that inflates findings marked disclosed, and says the bug shows up under the ledger’s disclosed filter. There is no technical write-up and no fix status. A screenshot of that filter is not an evidence exhibit. The author’s line is blunt: “I don’t know how anyone can read these numbers as anything but damning.” Read it as a claim about the public counters, not as proof about the unpublished 80.6 percent. 7,349 public findings out of 37,966 is the 19.4 percent. The other 80.6 percent is still inside Anthropic’s process.
Who validates Anthropic OSS Scanner reports now?
On 8 October 2026 Anthropic launched a self-service OSS Scanner and shifted validation to maintainers who opt in. The company said it had found more than 29,000 candidate vulnerabilities and had manually reviewed only about 6,000. It remained “bottlenecked on our human capacity to validate these findings,” and it had already sent nearly 5,000 unverified reports to maintainers who asked for everything in bulk.
About 6,000 manual reviews against more than 29,000 candidates leaves most of the pile untouched. The nearly 5,000 bulk sends went out unverified because maintainers requested them. A team that consumes those libraries inherits whatever queue the maintainer has not finished. OSS Scanner is the product response: a project requests a scan and triages the output itself. That is a different offer from the April line about Anthropic carrying responsible disclosure of high- and critical-severity findings.
Anthropic calls OSS Scanner output “fully model-generated, without human review or triage,” and says reports may be incorrect or invalid. Projects must opt in. The label moves the cost of a wrong report onto the maintainer. VulnCheck’s line is direct: “the triage bottleneck didn’t go away. It moved to the maintainers.” Opt-in means the project accepted a queue. It does not mean Anthropic finished validation.
Severity moved with the reports. In September 2026 VulnCheck found Claude rated 91.5 percent of ledger findings critical or high, while maintainers rated 51.3 percent that way. Do not inherit a model severity score. Recalculate it on the build you run, including preconditions and impact, and record the human who signed it. An AI security scanner built for your infrastructure is scoped and checked against your stack. Glasswing output belongs with other untrusted inbound advisories until a maintainer, or your own reviewer, confirms version, path, input, and impact.
What did curl do with 45 Anthropic vulnerability reports?
Curl maintainer Daniel Stenberg said the project received 45 Anthropic vulnerability reports, announced about three days before the VulnCheck piece. Anthropic rated one medium and the other 44 low. He is filing them as GitHub issues at a deliberate pace “to soften the impact for the team,” aiming to finish that week. Of the 34 published reports, 19 were marked not a vulnerability, 6 were confirmed low severity, and 9 await assessment. The false-positive rate on that set is 55.9 percent, and the confirmed share is 17.6 percent, all low severity. Eleven of the 45 were unpublished, so the rates cover that slice only.
Curl had already been through a smaller version of this in May 2026. Stenberg wrote that five Mythos findings reduced to one confirmed vulnerability. VulnCheck presents that five-to-one cut as the maintainer’s public account and does not re-test those five findings. Nineteen rejections out of 34 published reports is the cost of volunteer review time on this batch, spent to reach six low-severity confirmations.
VulnCheck calls publishing the low and medium reports as public issues a bold move and urges other maintainers to do the same. On this batch the model was not claiming any critical-severity findings. The April line about thousands of high- and critical-severity bugs and this curl batch can both be true, because they describe different populations. The disposition work is a secure code review: confirm, downgrade, or reject, with a human on the diff. Nineteen of 34 is the rejection rate a buyer can cite with the denominator attached. It is a rate for the published curl slice, not for all 37,966 ledger rows, and not for the 11 curl reports still unpublished.
What should a buyer require before a model report is a finding?
Buyers can conclude that the public 19.4 percent does not, on VulnCheck’s reading, support the 7 April claim of a fresh flow of high- and critical-severity fixes. They cannot say that about the other 80.6 percent, because those findings are not public. The 442 figure is public attribution assembled from CVE records, GHSAs, kernel commits, the ledger, and manual advisory checks. Rows still inside the unpublished 80.6 percent stay outside that credit count until they surface in one of those artifacts.
Five public items would change how far the November counters can travel: the true-positive denominator, a corrected disclosed filter, maintainer results for the non-public share, product and severity detail for CVE-2026-61500 and CVE-2026-26980, and a statement on whether 15 fixes in two months are early, late, or typical. The November write-up records no Anthropic reply that supplies the missing denominator, the filter fix, or product detail for the two KEV entries.
Before a model report enters a sprint, require four fields: affected version, reachable code path, triggering input, and security impact. A missing field stays in intake with a rejection reason. Track analyst minutes per report, duplicate rate, rejection reason, and time to fix. Publish the states and keep the words stable: candidate, rejected, duplicate, confirmed, fixed, disclosed. A claimed true-positive rate is unreadable if pending, duplicate, and withdrawn rows are undefined.
Do not let disclosed mean the model said so, or that Anthropic’s disclosed filter said so, while that filter is reported to over-count. If you opt in, cap weekly intake and name an owner. Define “not a vulnerability” in the issue so the next reader does not repeat the triage. A candidate ledger is not a pentest finding. AI-powered penetration testing should report impact a human operator reproduced. When discovery volume outruns fixes, fund an owner, a patch window, and columns that separate credit from confirmation.
FAQ
What is Anthropic Project Glasswing?
Project Glasswing is Anthropic's LLM-driven open-source vulnerability discovery program, launched on 7 April 2026. The launch said Anthropic had identified thousands of additional high- and critical-severity vulnerabilities and was working on responsible disclosure to maintainers. VulnCheck's November piece is the fifth public check of that claim.
How many CVEs has Anthropic Project Glasswing been credited with?
VulnCheck's tracker, started 17 April 2026, attributes 442 CVEs to Anthropic over about seven months, roughly 63 a month and about 0.5 percent of 93,450 CVEs published in 2026. Anthropic's ledger, as of 2 November, marks 217 findings fixed. The two counts answer different questions: public credit versus a fix flag.
What is Anthropic OSS Scanner?
OSS Scanner is Anthropic's opt-in, self-service scanner, launched 8 October 2026. Anthropic says the reports are fully model-generated, without human review or triage, and may be incorrect or invalid. At launch it cited more than 29,000 candidates, about 6,000 manual reviews, and nearly 5,000 unverified reports already sent in bulk.
Why did the curl project reject Anthropic vulnerability reports?
Of 34 published reports from a batch of 45, curl maintainer Daniel Stenberg's project marked 19 as not a vulnerability, confirmed 6 as low severity, and left 9 pending. Anthropic had scored the wider batch as 44 low and 1 medium. Eleven reports were still unpublished, so 55.9 percent false positives describes the published slice.
If model-generated dependency reports are already landing in your queue, LLM application penetration testing separates a reproduced boundary cross from a candidate no human has read.