VulnCheck says 217 of Anthropic's 37,966 findings are marked fixed; Anthropic's dashboard says 629 patched
VulnCheck says 217 of Anthropic's 37,966 findings are marked fixed. Anthropic's own dashboard, from the same 2 November snapshot, says 629 have been patched, and the two pages count different things.
By Parminder Kumar Sharma · LinkedIn · · 21 min read
217 fixed on one Anthropic page, 629 patched on another, both from the same 2 November snapshot
Two Anthropic pages carry the same snapshot time, 2026-11-02 19:30 UTC, and print two different fix counts. Its disclosure ledger shows a status filter, FIXED, with 217 entries. Its coordinated vulnerability disclosure dashboard says: "To our knowledge, 629 of these have been patched." The gap is 412 and the ratio is 2.9 (both derived). VulnCheck's post of 9 November 2026 and Infosecurity Magazine's report of 11 November 2026 use the 217. Neither prints the 629.
That is not proof of an error, because the two pages say they count different things, and this briefing sets out what each says. What the record does not establish is which number is the better measure of impact, how many of the fixes are in widely deployed software, how severe the fixed bugs were, or why the dashboard's count is nearly three times the ledger's. Anthropic's own dashboard calls patches a better indicator than true positives ("Another, more reliable one is the number of patches created", though it adds that patches are a lagging indicator), and VulnCheck's headline does not use that figure.
Read on 11 November 2026. The two Anthropic pages were saved as text by the lead from a browser tab and read here from those copies. No anthropic.com page was fetched for this briefing, so the dashboard's current figures, its About page glossary, its full ledger table and its archived snapshots were not checked, and anything that depends on them is marked as not verified. Figures labelled derived were computed here.
A naming point. VulnCheck and Infosecurity Magazine call the programme Project Glasswing. The two Anthropic pages saved for this briefing never use that word: they say Claude models, "including an early snapshot of Claude Mythos Preview", find vulnerabilities in open source software. This briefing uses VulnCheck's name only when reporting VulnCheck, and it does not establish that the dashboard and the ledger are, in Anthropic's terms, the Glasswing results.
What each number says it counts
The pages define their terms in places, and leave some undefined. The table sets each figure beside the page's own words and beside what the saved page does not say. The dashboard's long explanatory text is Anthropic's; the ledger page is saved as its header, filters and first 25 rows only.
Anthropic's figures, as the two saved pages (snapshot 2 November 2026, 19:30 UTC) describe them
| Figure and page | What the page says it counts | What the saved page does not say |
|---|---|---|
| 217, FIXED (ledger) | A status filter beside DISCLOSED, PRE DISCLOSURE, WITHDRAWN and DUPLICATE. Each entry shows its status once the report has been sent. | What FIXED means, or whether it counts every fix or only entries whose details are revealed. The About glossary was not saved. |
| 629, patched (dashboard) | "Patched upstream" counts reported vulnerabilities maintainers have fixed, "whether or not they replied to the report". The page says it does not guarantee that patches have been widely installed. | Which projects, which severities, or when. The headline sentence begins "To our knowledge". |
| 5,087, disclosed | The headline sentence says 5,087 vulnerabilities disclosed across 680 open source projects. The funnel labels the same figure as total reported to maintainers. | That all 5,087 were checked: the funnel puts 3,225 of them under "May contain false positives". |
| 7,349, ledger entries | Each entry is a SHA-3-512 hash of one finding's details. The dashboard describes the ledger as hash commitments for findings reported or planned. | How many entries are public findings. Before the maintainer is told, only the commitment hash and the commitment date are published. |
| 733, identifiers | 230 CVE records plus 503 GitHub Security Advisories; "a single finding may carry both". 230 plus 503 is 733 (derived check). | A count of distinct findings with an identifier. Because the sum adds the two lists, it counts identifiers, not findings. |
| 37,966, top of the funnel | The funnel's first figure. The page: "our models find candidate vulnerabilities, which we add to a list for human triage. This is the figure at the top." | That any of them is a true vulnerability. The 94.2% rate is of 6,376 reviewed findings, not of 37,966. |
One reading choice matters. The saved dashboard text prints each number before its label, which is easy to misread as label then number. Two checks agree with number then label: 6,009 divided by 6,376 is 94.2%, the printed rate (derived), and 1,862 plus 3,225 is 5,087, the printed total (derived). This site's earlier briefing on Anthropic's OSS Scanner read the dashboard's 2 October snapshot the same way. It means 6,376 is the number reviewed by outside firms, not a count of "candidates", and 37,966 is the candidate count.
The funnel: 5,087 disclosed, of which 3,225 had no independent check
Read in order, the dashboard says the models found 37,966 candidates. Six outside firms reviewed 6,376 of them (16.8%, derived) and confirmed 6,009, which is the 94.2%. 1,862 findings were reported to maintainers after that review. The other 3,225 were "Reported direct to maintainers by Anthropic", with the page's own warning: "May contain false positives". The direct share is 63.4% of the 5,087 (derived). The dashboard says direct disclosure also happens "when maintainers specifically request" un-triaged findings.
The true positive rate is defined on the page, and the definition bears on VulnCheck's question about duplicates. Anthropic says the rate counts findings the firms confirmed as valid, as a share of those manually reviewed, and that it "includes real bugs that we later discover have already been reported" and "won't fix" findings. So by the page's own definition, duplicates found later are inside the 94.2%. The page does not say how withdrawn findings are treated. It also calls the rate "one proxy for impact" and says it should be read with caution, because it relies on the firms rather than on maintainers.
The dashboard also says why the disclosed figure trails the finding figure: independent human triage "is the rate limiting step", and many bugs confirmed as real are not yet reported "due to capacity limitations". If the 6,009 confirmed and the 1,862 reported after triage are compared, 4,147 confirmed findings are not in the triaged route (derived; the page does not print this number or say all are eligible).
The dashboard against its own earlier snapshot. This site's earlier briefing on OSS Scanner (read it here) saved the dashboard's machine readable record from 2 October 2026 (revision 35). Set side by side, the candidate count grew and the reported counts fell:
Dashboard funnel, 2 October 2026 snapshot (saved for the earlier briefing) against the 2 November 2026 snapshot; the change column is derived
| Funnel step | 2 Oct 2026 | 2 Nov 2026 | Change |
|---|---|---|---|
| Candidates | 29,439 | 37,966 | +8,527 |
| Reviewed by outside firms | 6,123 | 6,376 | +253 |
| Confirmed valid (rate) | 5,674 (92.7%) | 6,009 (94.2%) | +335 |
| Reported after triage | 1,333 | 1,862 | +529 |
| Reported direct | 4,824 | 3,225 | -1,599 |
| Total reported to maintainers | 6,157 | 5,087 | -1,070 |
| Acknowledged by maintainer | 5,103 | 3,635 | -1,468 |
| Patched upstream | 516 | 629 | +113 |
| CVEs and GHSAs assigned | 584 | 733 | +149 |
The total reported to maintainers is lower on 2 November than on 2 October by 1,070, and the direct reports are lower by 1,599, while the candidates grew by 8,527. The saved pages do not say why. A reclassification of some direct reports, for example into duplicates or withdrawals, would fit the arithmetic, but that is this site's conjecture and neither page says it. It matters here because the 5,087 is the base for the 629: a base that moved by 1,070 in a month is a base to name when quoting a patched share.
The ledger: 7,349 entries, and five statuses that add to 7,570
The ledger header prints 7,349 of 7,349 entries and five status counts: DISCLOSED 5,087, FIXED 217, PRE DISCLOSURE 193, WITHDRAWN 419 and DUPLICATE 1,654. Added together they come to 7,570, which is 221 more than the entries (derived). An entry has one status, so either the filters overlap or a count is off; the ledger text does not say which. Leave FIXED out and the four others sum to 7,353, still four over.
VulnCheck's figure shows what it did about this. It strikes out 5,087 and prints 4,866 as disclosed, and says in its post: "There appears to be a bug on Anthropic's ledger that inflates the number of findings marked as disclosed." With 4,866 the five groups sum to exactly 7,349 (derived), and 5,087 minus 4,866 is 221, which is 217 plus 4. That reads as if VulnCheck treats the ledger's disclosed filter as already containing the fixed entries, with four more unexplained. The post does not give that arithmetic, so this is a reading. The consequence is that the dashboard headline, which is also 5,087, carries the same figure VulnCheck says is inflated; whether the dashboard has the same overlap is not stated.
A lead on what FIXED may be. The 2 October dashboard record saved earlier carries a field named total_fixed, 218, and a block named revealed with 235 entries of which 218 fixed, next to a separate fixed_in_response, 516, which matches that day's "Patched upstream". If the ledger's FIXED status is the first and the dashboard's patched figure is the second, FIXED would count fixes among entries whose details have been revealed, and patched would count every fix on a reported finding. That would explain a gap of this kind, but the 2 November record was not saved, the mapping is not stated anywhere, and the lead stays a lead. Note also that 218 on 2 October and 217 on 2 November is a fall of one.
The ledger text does explain why many entries show little. Before the maintainer is told, "only the commitment hash and the commitment date are published"; the identifier, project and bug class appear only when the disclosure window closes. The saved pages do not give the window's length. The dashboard lists CVE records in pages of ten ("Page 1 of 10") and GHSAs in seven pages, so at most about 100 CVE records are listed publicly of the 230 assigned (derived from the page count); the page says unlisted identifiers belong to findings whose window has not closed or that the assigning authority has not published.
What VulnCheck says, how it counted, and where its post differs from itself
VulnCheck's post, dated 9 November 2026, is a seven month update on its tracking of Anthropic's vulnerability work. Its method, in its words, is a tracker: "every vulnerability attributable to Anthropic using CVE records, GHSAs, Linux kernel commits, and Anthropic's own ledger", with manual audits of advisories, begun on 17 April. It does not name the dashboard. Its headline claims are these, all VulnCheck's:
VulnCheck's key claims, as its post prints them, with what each rests on and an alternative reading from the saved Anthropic pages
| VulnCheck's claim | What it rests on | What Anthropic's pages say |
|---|---|---|
| 442 CVEs attributable to Anthropic, 0.5% of 93,450 CVEs published this year (0.47%, derived) | Its own tracker repository, which the post links. This briefing did not read the repository. | The dashboard counts 230 CVE records. 442 is 1.92 times 230 (derived), counted by different rules. |
| Only 7,349 of 37,966 findings (19.4%) have been made public | The ledger's entry count against the dashboard's top figure. | 37,966 are candidates for human triage. Entries carry hash only until the report is sent, so entries are not the same as public findings. |
| 217 marked fixed | The ledger's FIXED status. | 629 patched upstream, to Anthropic's knowledge, on the dashboard. |
| 1,654 duplicates (4.4%), nearly eight times the fixes; 419 withdrawn | The ledger's DUPLICATE and WITHDRAWN statuses. 1,654 divided by 217 is 7.6 (derived); the post's body says 7.5x. | The true positive rate includes duplicates found later and won't fix findings. |
| The 94.2% true positive rate raises questions about duplicates and withdrawn findings | The post's reading: it calls the rate odd and likely lacking a feedback loop. | Duplicates are counted inside the rate by definition; withdrawn findings are not addressed on the saved page. |
| Only 2 of 442 CVEs exploited in the wild | VulnCheck KEV, which this briefing did not read. | Not stated. The two are named by VulnCheck: CVE-2026-61500 and CVE-2026-26980. |
The post's characterisation is opinion and is VulnCheck's: its title calls the findings "underwhelming", it calls the ledger update "a staggering failure" in one place and says of the numbers that it does not know how anyone can read them as anything but damning. It closes: "They still don't reconcile." Those words are VulnCheck's judgement, not a measurement.
Between 8 September and 2 November, VulnCheck says, the picture moved as follows. The table copies its cells; the last column is as printed.
VulnCheck's own comparison table, copied cell by cell from its post of 9 November 2026
| Measure | 8 September | 2 November | Change |
|---|---|---|---|
| Total findings | 26,153 | 37,966 | +11,813 |
| Made public | 2,736 (10.5%) | 7,349 (19.4%) | +4,613 |
| Fixed | 202 | 217 | +15 |
| Withdrawn | 245 | 419 | +174 |
| Duplicates | 2 | 1,670 | +1,668 |
| Claimed true positive rate | 91.4% | 94.2% | +2.8 pts |
Three places where the post differs from itself. First, duplicates are 1,670 in the table, 1,654 in the text and its figure, and 1,654 on the ledger (a difference of 16 that the post does not explain). Second, the ratio of duplicates to fixes is "nearly eight times" in the key takeaways, 7.5x in the body and 7.6 by division. Third, the tracker screenshot in the post shows 442 CVEs in its header while its two pie charts total 446 vulnerabilities; the severity bands in that screenshot (critical 50, high 189, medium 143, low 36, unscored 28) add to 446 (derived). The post's "seven months" also runs from the 7 April launch it cites, while its tracker begins on 17 April.
The two CVEs VulnCheck says are exploited, and what the CVE records say
VulnCheck names two Anthropic attributable CVEs as exploited in the wild: CVE-2026-61500 and CVE-2026-26980. The post prints the identifiers only. This briefing read both CVE records at CVE.org; the facts below are from those records, not from VulnCheck.
The two CVEs, from their CVE.org records (read 11 November 2026); the Anthropic row for the second is from the dashboard page
| Item | CVE-2026-61500 | CVE-2026-26980 |
|---|---|---|
| Product | Rejetto HFS | Ghost, Content API |
| Weakness (record title) | Session forgery via predictable signing key | SQL injection |
| Affected, fixed | 3.0.0 up to but not including 3.2.1 | 3.24.0 up to but not including 6.19.1; fixed in 6.19.1 |
| Record published | 13 July 2026; record assigned by VulnCheck | 20 February 2026; record assigned by GitHub |
| CVSS 3.1 (assigned by the record's CNA) | 9.8 critical | 9.4 critical |
| Anthropic link | Record credit line: an outside security firm "in collaboration with Claude and Anthropic Research" | Not in the record; the dashboard lists it as ANT-2026-69D8H6RP, critical, "CVSS 9.1" |
Four things follow, all derived from the records and the feed. The second CVE's record was published on 20 February 2026, before the 7 April launch VulnCheck cites, so "attributable to Anthropic" is not the same as a Glasswing result. The dashboard gives the Ghost finding a CVSS of 9.1 where the record's assigner gives 9.4, a difference between two scores for one flaw. VulnCheck's own record assigns the first CVE, so one of its two exploited examples sits in a record VulnCheck wrote; that is a fact about the record, not a charge. And neither identifier is in the CISA Known Exploited Vulnerabilities catalogue as of its feed dated 10 November 2026, 19:03 UTC (1,779 entries), although VulnCheck says it used its own KEV list, which is a different list. This site covered the first of the two earlier: CVE-2026-61500 was fixed on 13 July, with exploitation reported 80 days later.
The curl reports: a different population from the 94.2%
VulnCheck's post adds a maintainer data point. It says curl's maintainer announced three days earlier that the project had received a batch of 45 vulnerability reports from Anthropic, rated one medium and the rest low, and that of 34 published so far, 19 were marked not a vulnerability, 6 confirmed as low severity and 9 await assessment (19 plus 6 plus 9 is 34; 55.9%, 17.6% and 26.5% derived). VulnCheck calls it "a far cry" from the 94.2%. This briefing did not read curl's own pages, so the counts rest on VulnCheck, with Infosecurity Magazine repeating 19 and six.
The comparison is not like for like. VulnCheck's heading frames the batch as a test of Anthropic's OSS Scanner, whose reports the post quotes Anthropic as calling "fully model-generated, without human review or triage". The 94.2% describes findings that outside firms reviewed. The dashboard's direct reports, the nearest equivalent, carry no rate: the page prints a false positive warning instead. This site's earlier briefing on the OSS Scanner and its unreviewed reports covers that design. The curl numbers show what one maintainer found in one batch; they do not measure the 217, the 629 or the 94.2%.
Where VulnCheck, Infosecurity Magazine and Anthropic's pages differ
Infosecurity Magazine's article, with a published time of 16:30 UTC on 11 November 2026, is a secondary: it reports VulnCheck's analysis and adds one outside fact, that Anthropic had not responded by publication. Its wording differs from the post and from the Anthropic pages in several places, and the differences are kept here, not resolved.
How the three sources word the same things; Infosecurity Magazine's figures are printed without thousands separators as in its text
| Item | VulnCheck post, 9 Nov | Infosecurity Magazine, 11 Nov | Anthropic pages, 2 Nov |
|---|---|---|---|
| What 37,966 is | "Anthropic's 37,966 reported findings" | Glasswing "has enabled researchers to find 37,966" | Candidates found by the models, for human triage |
| 7,349 | Made public, 19.4% | "listed in the ledger", 19.4% | Ledger entries; some are hash only |
| Disclosed | 4,866 in its figure, 5,087 struck out | 4866 reported to maintainers, almost 13% | 5,087 |
| Fixed as a share | 0.5% in its figure | "less than 0.5%"; 217 of 37,966 is 0.57% (derived) | No share printed |
| Duplicates | 1,654 in text and figure, 1,670 in its table | 1654, and 1668 added since September | 1,654 |
| 5,276 | Not printed in the post text | "validated as real unique vulnerabilities" | Not stated. 5,276 is 7,349 minus 1,654 minus 419 (derived), and includes the 193 pre-disclosure entries |
| The 629 | Not mentioned | Not mentioned | Printed as patched upstream |
| curl | 19 of 34 not a vulnerability, 6 low, 9 awaiting | 34 analysed "using Anthropic OSS Scanner and other tools"; 19 not vulnerabilities, six low | Not applicable |
| Anthropic's response | None in the post | Contacted; "had not responded by the time of publication" | The pages are a 2 November snapshot, a week before the post |
Two differences deserve a plain sentence. Infosecurity Magazine's "validated" for 5,276 is stronger than anything the ledger says: the 5,276 includes 193 entries still before notification, and the saved ledger page says nothing about validation. And the article's "less than 0.5%" does not match the division, which gives 0.57%; VulnCheck's own figure rounds to .5%.
What is stated, and what is not
Questions a reader would ask of the two fix counts, and what the saved pages state
| Question | What the pages state | Not stated |
|---|---|---|
| Do the 217 and 629 count the same thing? | Each page defines its own term in part: FIXED is a ledger status; patched upstream means a maintainer released a fix. | A reconciliation of the two, or the definition of FIXED. |
| How many fixes are in widely deployed software? | The page warns that patches may not be widely installed. | Any count by project, popularity or deployment. |
| How severe are the fixed bugs? | The dashboard prints Claude against firm severity agreement, n = 1868 (the funnel says 1,862 reported after triage; the page does not explain the six): 88.0% exact, 97.9% within one band. VulnCheck's tracker screenshot shows CVSS bands for its 446 counted CVEs. | The severity of the 217 or the 629. |
| How many of the 37,966 are true vulnerabilities? | 94.2% of 6,376 reviewed were confirmed, a rate that includes duplicates found later and won't fix. | A rate for the other 31,590 unreviewed candidates (37,966 minus 6,376, derived), or for the 3,225 direct reports. |
| What do maintainers think? | Via VulnCheck only: one maintainer's view of 34 OSS Scanner reports. | Any maintainer statement read by this briefing. |
| How does VulnCheck attribute a CVE to Anthropic? | Its post names its inputs: CVE records, GHSAs, kernel commits, the ledger, manual advisory audits. | The rule for deciding attribution. The linked repository was not read. |
| Do the disclosure windows explain the low fixed count? | Details appear only when the window closes; hash and date come first. | The window's length, or how many entries are inside one. |
| Does Anthropic dispute VulnCheck? | Nothing found. Infosecurity Magazine says Anthropic had not responded by publication. | Any Anthropic statement on the post. |
What to do with an AI found vulnerability count
These steps are this site's suggestions, not advice from VulnCheck or Anthropic.
Take this with you
In order, for a reader who relies on AI found vulnerability counts
- Ask which stage a headline number counts: candidate, reviewed, confirmed, reported, acknowledged, patched or given an identifier, and write the stage beside the figure.
- Prefer fixed counts and assigned identifiers over raw findings, and confirm a fix in the project's own release notes or advisory before counting it.
- Read a vendor dashboard and an independent tracker side by side and record the snapshot time of each. Anthropic's was 2 November and VulnCheck's post was 9 November, a week apart.
- Treat a true positive rate as relative to its denominator, and ask whether duplicates, won't fix and withdrawn findings are inside it. Anthropic's page says duplicates found later and won't fix are.
- If you maintain open source code, plan triage capacity for unreviewed reports: 3,225 of the dashboard's 5,087 were reported direct with a page warning that they may contain false positives.
- If you act on an AI found CVE, check its record at CVE.org and your exploitation source; neither CVE VulnCheck names is in the CISA catalogue as of its 10 November feed.
- Re-read both pages before quoting any figure from this briefing, because the next dashboard update can change every number here.
Technical details
Neither source prints exploit code, and this briefing prints none. What a reader can use is the exact labels and counters the pages print, so that two snapshots can be compared label by label. A defender or analyst who follows these counts can copy the lines below into a spreadsheet and diff them against the next snapshot.
The first block is the dashboard funnel exactly as the saved text prints it, one item per line, number before label. Check the pairing against the arithmetic in the earlier section before relying on it.
DISCOVERED
TRIAGED
37,966 findings
Candidates
6,376 findings
Reviewed by external security firms
6,009 findings
Confirmed valid
94.2% true positives of 6,376
1,862 findings
Reported to maintainers
3,225 findings
Reported direct to maintainers by Anthropic
May contain false positives
DISCLOSED
REMEDIATED
5,087 findings
Total reported to maintainers
3,635 findings
Acknowledged by maintainer
629 findings
Patched upstream
733 advisories
CVEs and GHSAs assigned
Counts as of November 2, 2026
The second block is the ledger page's header counters, filter values and column names. A status filter is the term a reader should quote, because the saved page does not define FIXED.
MATCHING
7,349
of 7,349 entries
DISCLOSED
5,087
FIXED
217
PRE DISCLOSURE
193
WITHDRAWN
419
DUPLICATE
1,654
STATUS
All statuses
disclosed
duplicate
fixed
pre-disclosure
withdrawn
ASSESSED BY
Any source
Claude-assessed
Firm-assessed
Maintainer-assessed
SEVERITY
Any severity
Critical
High
Medium
Low
DATE COMMITTED
HASH / IDENTIFIER
PROJECT
BUG CLASS
SEVERITY ASSESSMENTS
STATUS
The page names its own provenance lines. The manifest hash and the dated archive are how a reader can verify a figure against the page live on a date; this briefing did not fetch them.
Snapshot generated at 2026-11-02 19:30 UTC
Dates and times are in UTC
Manifest hash 25d811e7b56632ef5bbb23eb…
Archived snapshots remain available at their dated paths so any figure can be verified against the page that was live on that date
payload.json
VulnCheck's own comparison table, as a defender would paste it into a sheet. Row labels are VulnCheck's; it prints 1,670 for duplicates on 2 November where the ledger shows 1,654.
Measure | 8 September | 2 November | Change
Total findings | 26,153 | 37,966 | +11,813
Made public | 2,736 (10.5%) | 7,349 (19.4%) | +4,613
Fixed | 202 | 217 | +15
Withdrawn | 245 | 419 | +174
Duplicates | 2 | 1,670 | +1,668
Claimed true positive rate | 91.4% | 94.2% | +2.8 pts
The two CVEs VulnCheck says are exploited, with version ranges and score vectors from their CVE records. A defender running either product compares the affected range with the installed version; the HFS range ends before 3.2.1 and the Ghost range before 6.19.1.
CVE-2026-61500 (record assigned by VulnCheck, published 2026-07-13)
Title: Rejetto HFS < 3.2.1 Session Forgery via Predictable Signing Key
Affected: 3.0.0 up to but not including 3.2.1
CVSS 3.1 (record's CNA): 9.8 CRITICAL CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
Credit line: ... in collaboration with Claude and Anthropic Research
CVE-2026-26980 (record assigned by GitHub_M, published 2026-02-20)
Title: Ghost has a SQL Injection in its Content API
Affected: >= 3.24.0, < 6.19.1 (fixed in 6.19.1)
CVSS 3.1 (record's CNA): 9.4 CRITICAL CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:L
Anthropic dashboard entry: ANT-2026-69D8H6RP, critical
Dashboard title: Critical Vulnerability (CVSS 9.1) : SQLi from unauthenticated user; arbitrary DB reads
The question the two counts leave open
Anthropic's dashboard says patches are the more reliable measure of impact, and its own two pages, from one snapshot, print 217 for fixed and 629 for patched, with a reported total that fell by 1,070 in a month. VulnCheck says the numbers do not reconcile and prints a disclosed count that Anthropic's headline does not. If the measure that matters is a fix in a project's own release, which of those numbers is it, and who will publish the list of fixes it counts?
Key facts
Sources
- PrimaryBlog post dated 9 November 2026, Glasswing seven month update: 442 attributable CVEs, ledger counts, curl reports, comparison table and figures. Read in full; the figure and tracker screenshots savedVulnCheckaccessed 2026-11-11
- PrimaryCoordinated vulnerability disclosure dashboard, last updated 2026-11-02 19:30 UTC: 5,087 disclosed, 629 patched, funnel and definitions. Read from text saved by the lead; not fetched by this briefingAnthropicaccessed 2026-11-11
- PrimaryDisclosure ledger page, snapshot 2026-11-02 19:30 UTC: header counts 7,349 entries and five status counts, filters, first 25 rows. Read from text saved by the lead; the full table was not savedAnthropicaccessed 2026-11-11
- PrimaryCVE record for CVE-2026-61500: Rejetto HFS, version range, CVSS, credit lineCVE Program (CVE.org)accessed 2026-11-11
- PrimaryCVE record for CVE-2026-26980: Ghost SQL injection, version range, CVSS, publication dateCVE Program (CVE.org)accessed 2026-11-11
- PrimaryKnown Exploited Vulnerabilities feed dated 2026-11-10: checked that neither CVE is listedCISAaccessed 2026-11-11
- Reported byNews report of 11 November 2026 on VulnCheck's analysis, with the 5,276 and 4,866 figures and a note that Anthropic had not responded. Secondary; read in fullInfosecurity Magazineaccessed 2026-11-11
