# Anthropic's free security scans give curl 45 reports to check

**Summary:** Anthropic's free security scans require maintainer review. curl's reports show the workload, and Glasswing's 217 fixed ledger entries are not its full reported patch count.

- Canonical: https://markhuang.ai/news/anthropic-oss-scanner-curl-review-queue
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-11-10
- Section: News
- Tags: Anthropic, Project Glasswing, OSS Scanner, curl, AI Security, Open Source
- Source: [VulnCheck](https://www.vulncheck.com/blog/anthropic-glasswing-november-update)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![A magnifying lens reveals a small crack in one embossed bug card beside a queue of cards under an optical scanner](https://cdn.markhuang.ai/news/anthropic-oss-scanner-curl-review-queue/hero.webp)

*Conceptual illustration: automated scans produce candidate findings; review determines which ones deserve a security fix. Generated with imagegen.*

One curl reviewer agreed that an Anthropic report described a valid, low-severity vulnerability. He also called it "an essay of thousands of words" about something simple. In the [same public discussion](https://github.com/curl/curl/issues/23759#issuecomment-6681099846), curl maintainer Daniel Stenberg confirmed that the team had 45 reports to work through, publishing them gradually. Even a useful finding can arrive with a lot of reading.

Anthropic's [OSS Scanner, launched October 8, 2026](https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source), gives eligible open-source projects recurring security scans at no cost. Maintainers can receive findings faster by opting into reports generated entirely by models, without human review. That makes another source of security analysis available to projects willing to do the checking themselves.

[Patrick Garrity's November 9 Glasswing update for VulnCheck](https://www.vulncheck.com/blog/anthropic-glasswing-november-update) examines the work behind those reports. The maintainer comments make the workload hard to dismiss. Some headline counts need more care: a candidate finding, a confirmed vulnerability, and a patch describe different stages of the job.

## A real bug still needs a readable report

The [accepted curl report](https://github.com/curl/curl/issues/23759) concerned a request to require TLS 1.3, a version of the protocol used to secure network connections. It described how curl's Windows Schannel configuration could instead allow TLS 1.2 under specified conditions. A reviewer accepted the issue as low severity and pointed to a simpler explanation and fix. The report itself acknowledged that its reproduction used a stand-in for Windows APIs on Linux, rather than a real Windows installation. I have not reproduced it.

Another report shows why reproducing behavior is only part of validation. It claimed that mixed-case cookie names could bypass protections. [Stenberg rejected its premise](https://github.com/curl/curl/issues/23773#issuecomment-6680111094), pointing to a specification's case-sensitive matching rule. The scanner had criticized curl for making that exact match. The maintainer had to check what the software was supposed to do before deciding whether the alleged failure was a bug.

Garrity's November 9 snapshot counted 34 published curl reports: 19 marked as not vulnerabilities, six confirmed as low severity, and nine awaiting assessment. Those are his interim counts; assessment continued as the team published more of the 45-report batch. A finding rejected as a security vulnerability may still identify an ordinary bug worth fixing. The tally also cannot establish how the scanner performs across other projects.

## What the 94.2% accuracy figure measures

VulnCheck contrasts curl's experience with Anthropic's claimed 94.2% true-positive rate. The [November 2 disclosure dashboard](https://red.anthropic.com/2026/cvd/index.html) explains the denominator: external security firms reviewed 6,376 candidate findings and confirmed 6,009 as valid. That percentage measures the reviewed subset. It does not describe every model-generated report delivered to maintainers.

Anthropic also says this measure includes real bugs later found to be duplicates and findings a maintainer may decline to fix, including those outside the project's threat model. A threat model defines which attacks and conditions the project intends to defend against. External reviewers and maintainers can therefore disagree about whether an issue warrants action without necessarily disagreeing about the code's behavior.

The October launch announcement says Anthropic will continue human-verified coordinated disclosures alongside the optional scanner. The two routes assign the validation work differently:

```mermaid
flowchart TD
    A["AI-generated candidate finding"] --> B["Human-reviewed disclosure route"]
    A --> C["Opt-in OSS Scanner route"]
    B --> D["Security researchers validate the finding"]
    D --> E["Maintainer assesses the report"]
    C --> F["Model-generated report sent without human review"]
    F --> E
    E --> G["Accepted findings can lead to a patch"]
```

curl's mostly low-severity batch and the externally reviewed disclosure subset are different samples. Their percentages cannot establish the scanner's overall accuracy. A maintainer needs evidence about the reports they will actually receive.

## The ledger's 217 fixes are not the whole patch count

VulnCheck reports 37,966 findings and 217 marked fixed in the November 2 ledger, with only 15 more fixed entries than in September. The [public ledger](https://red.anthropic.com/2026/cvd/ledger/) does show 217 entries with that status. But Anthropic's overview for the same snapshot separately reports 629 vulnerabilities patched upstream.

These are different views of the program. The public ledger table lists 7,349 entries and explains how details become visible through disclosure. The overview tracks broader progress, including patches maintainers have released. I have not reconciled the individual records behind the two patch totals, but 217 alone is an incomplete account of Anthropic's reported remediation.

Anthropic cautions that an upstream patch does not guarantee widespread installation. Still, the broader patch count belongs beside the critique. The review burden deserves scrutiny alongside the fixes the program reports.

## Make the reports fit the project's review capacity

For maintainers considering enrollment, the [OSS Scanner documentation](https://github.com/anthropics/oss-scanner) offers useful controls. Projects can supply a threat model, report guidelines, and patch guidelines. They can also pause reports. Given curl's experience, I would use those guidelines to ask for a concise explanation of the violated security rule and reproduction steps that make assumptions explicit.

My preferred measure is accepted security fixes relative to the review time needed to reach them. That is an evaluation I recommend, not a result measured here. Record useful ordinary bug fixes separately, so a security rejection does not erase their value. The verification problem in my [earlier post on AI-assisted WordPress vulnerability research](https://markhuang.ai/news/gpt-5-6-wordpress-rce-human-bottleneck) now appears in a queue that maintainers must manage.

The opportunity is practical: projects can get another search for flaws without paying for the scan. Whether that saves the team time depends on what arrives and how readily someone can verify it. curl's public reviews let other maintainers examine that work before signing up.
