Project Glasswing results tracker: findings, fixes, CVEs and the true positive claim, dated

This page tracks one program's claims against the evidence. Project Glasswing is Anthropic's initiative to find and fix vulnerabilities in critical software using its frontier models. It launched with a strong claim and no public ledger. Seven months later there is a ledger, an independent tracker and a maintainer whose project received a batch of reports. We log each dated data point here and update the page through corrections, not new articles.

The claim at launch

Anthropic's Glasswing page records the launch on April 7, 2026: a new initiative bringing together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks to secure critical software. It described an unreleased frontier model, Claude Mythos Preview, as capable of surpassing all but the most skilled humans at finding and exploiting software vulnerabilities, and said it had found thousands of high-severity vulnerabilities, including some in every major operating system and web browser.

According to VulnCheck, the launch text also said: "We have identified thousands of additional high- and critical-severity vulnerabilities that we are working on responsibly disclosing to open source maintainers." On October 6, 2026, Anthropic's page says it expanded its Cyber Verification Program into three access tiers, with existing Glasswing members moving to a Specialized Access tier.

The dated log

Each row below is a dated claim or measurement. Sources are named, and the source type matters: the ledger is Anthropic's own, while VulnCheck's counts and the curl numbers are independent.

• April 7, 2026: Glasswing launches (Anthropic).
• April 17, 2026: VulnCheck begins tracking CVEs attributable to Anthropic in an open repository (VulnCheck).
• May 2026: Anthropic's ledger launches reporting 23,019 findings and a 90.8% true positive rate (Anthropic ledger, via VulnCheck).
• September 8, 2026: Ledger shows 26,153 findings, 2,736 public (10.5%), 202 fixed, 245 withdrawn, 2 duplicates and a 91.4% claimed true positive rate (via VulnCheck).
• October 8, 2026: Anthropic's OSS Scanner post says it found more than 29,000 candidate vulnerabilities but manually reviewed only about 6,000, and had sent nearly 5,000 unverified reports to maintainers who asked for everything in bulk (Anthropic, as quoted by VulnCheck).
• November 2, 2026: Ledger shows 37,966 findings, 7,349 public (19.4%), 217 fixed, 419 withdrawn, duplicates in the range of 1,650 to 1,670 and a claimed 94.2% true positive rate (via VulnCheck).
• November 9, 2026: VulnCheck publishes its seven-month review.

What the independent numbers say

VulnCheck's November 9 review, by Patrick Garrity, counts 442 CVEs attributable to Anthropic since it began tracking, an average of about 63 a month and roughly 0.5% of the 93,450 CVEs published so far in 2026. Only two, CVE-2026-61500 and CVE-2026-26980, appear in VulnCheck's known-exploited list, which it puts at 0.45%.

Between September 8 and November 2, VulnCheck says the ledger added 11,813 findings and just 15 fixes. Its text gives the duplicate count as 1,654 (4.4%) while its table lists 1,670, a small internal discrepancy we note rather than resolve. Either way, the duplicate figure is several times the fixed figure.

VulnCheck also flags an apparent bug in Anthropic's ledger that inflates the number of findings marked as disclosed, and argues the claimed true positive rate rose while withdrawals and duplicates exceeded fixes, suggesting a feedback problem in how triage rates are counted. That reading is VulnCheck's inference. Anthropic has not, in the sources we opened, explained the accounting.

The maintainer test

The sharpest data point comes from curl. VulnCheck says Daniel Stenberg, who maintains the project, announced a batch of 45 vulnerability reports from Anthropic, with only one rated medium severity and the rest low. Of 34 published so far, 19 were marked not a vulnerability, 6 confirmed low severity and 9 still awaiting assessment. VulnCheck computes a 55.9% false-positive rate on those 34 and notes only 17.6% were confirmed, all low severity. These are partial results for one project.

Anthropic's own OSS Scanner post, quoted by VulnCheck, says its reports are "fully model-generated, without human review or triage" and that "it is possible reports will be incorrect or invalid." In other words, validation work moved to maintainers.

What it does and does not tell us

VulnCheck is a vulnerability-intelligence vendor with an interest in being the neutral scorekeeper, and its tone is critical. Its own conclusion is that AI is a valuable tool for finding vulnerabilities. What it disputes is the launch framing. Our reading: the program produces a large pile of candidate findings and a small pile of confirmed, fixed, exploited ones, and the gap is the story.

For the capability side, see our coverage of the Claude Mythos 5.2 system card and our tracker of restricted-access cyber AI models.

Why the definitions matter

A "finding" in Anthropic's ledger is a candidate vulnerability reported by a model, not a CVE assigned by a maintainer or vendor. A "fix" means a patch has landed. A "true positive" is a judgment by someone at triage. Each step filters the last, so comparing findings to fixes mixes categories, and VulnCheck's argument is that the filter is tightening more slowly than the intake grows. Anthropic's own October statement that it remains "bottlenecked on our human capacity to validate these findings" is consistent with that reading.

How we will update this page

We will add a row whenever Anthropic updates its ledger, VulnCheck publishes new counts, or a maintainer publishes results. We will mark each row with who produced it. We will not average claims with measurements.

Our position: the number to watch is fixes, not findings, and the second number is how many findings maintainers confirm as real. A program that finds 38,000 candidates and fixes 217 has a triage bottleneck, which Anthropic itself said in October. Until fixes catch up, treat the true positive rate as a vendor metric.