Cheng Qian, from Aofeisi

Qubit AI | WeChat Official Account QbitAI

Recently, the issue of AI finding bugs has become quite fragmented.

On one side, AI, once it steps in, reviews everything, first uncovering a major security vulnerability in PHP's core , then exposing 948 Linux kernel vulnerabilities ……

On the other side, the open-source community is almost overwhelmed: you found the vulnerabilities, but where's the fix!

Yesterday, the cybersecurity company VulnCheck released a report on Anthropic's Glasswing Project .

The report shows that seven months after the launch of the Glasswing Project, Claude has identified a total of 37,966 candidate vulnerabilities .

However, as of now, only 217 have been publicly recorded as fixed.

VulnCheck couldn't help but question: has the legendary cybersecurity era "Mythos" really arrived?

38,000 vulnerabilities, how many have been actually fixed?

Let's first recall the "Glasswing Project" that Anthropic launched in April this year with much fanfare.

The project aims to use Claude to extensively discover security vulnerabilities in critical software and open-source projects, collaborating with tech companies and open-source maintainers to promote vulnerability fixes, eliminating potential security risks before they are exploited.

The name carries two implications.

On one hand, security vulnerabilities are like glasswing butterflies, hidden in plain sight for decades without being noticed; on the other hand, transparency itself is a form of protection, and Anthropic hopes that through more transparent security research and vulnerability disclosure mechanisms, software systems can gain better protection.

△ Glasswing Butterfly, scientific name: Wide-band Black-winged Moth | Image source: wikipedia

At the launch, Anthropic claimed to provide 100 million model usage credits and 4 million dollars in open-source security donations, and also announced that their cybersecurity top-tier model Claude Mythos Preview had discovered thousands of high-risk and critical vulnerabilities. Soon after, Apple, Google, Microsoft, AWS, and NVIDIA joined in.

But is that really the case?

According to VulnCheck's continuous tracking data, as of early November, Anthropic's candidate vulnerabilities have reached 37,966, but in the vulnerability disclosure records, only 217 are marked as fixed. Meanwhile, the number of invalid data is still piling up, with 419 vulnerabilities withdrawn and 1,654 duplicates.

△ Image source: VulnCheck

The gap is still growing.

Within the past two months alone, the number of candidate vulnerabilities has nearly doubled by 12,000, but the number of fixed records has only slightly increased by 15.

In addition, VulnCheck tracked 442 CVEs attributed to Anthropic, accounting for only 0.5% of the total number of publicly disclosed CVEs worldwide in 2026.

According to data disclosed by Anthropic's official, as of November 2, out of 37,966 candidate findings, 6,376 have been reviewed by external security experts, with 6,009 findings being valid. Among these, 5,087 have been reported to software maintainers, and 629 have already received upstream patches.

Despite the slight differences in the number of fixes due to varying data metrics, there is one thing both sides have noticed.

That is, the intensity of AI in finding vulnerabilities is overwhelming for manual review.

Anthropic has already recognized this issue. The official also admits that the severe bottleneck of manual verification greatly limits the efficiency of vulnerability review and disclosure.

So on October 8, it launched a service called OSS Scanner.

Open-source project maintainers can voluntarily sign up to have Claude scan their code for free. This way, the vulnerability reports generated by the model can be directly sent to the maintainers without waiting for Anthropic to manually review them.

Open-source maintainers can't fix all the issues

AI is forcing us to immediately settle a 40-year-old technical debt.

Behind this data gap, the most affected are the open-source maintainers. For example, VulnCheck cited the network data transmission tool curl.

In early November, curl maintainer Daniel Stenberg received another large batch of vulnerability reports from Anthropic,45 vulnerability reports.

And before that, Daniel wrote in his blog that the core maintainers were already overwhelmed by the continuous influx of reports.

Our goal is to not drown. (We aim to not drown.)

To deal with so many reports, Daniel announced a change to curl's vulnerability handling rules. Low and medium severity vulnerabilities are now publicly discussed on GitHub,allowing more community developers to help with the review.

Currently, among the 34 publicly disclosed reports for curl, 19 were determined not to be security vulnerabilities, and 6 were confirmed as low severity.

VulnCheck's report subtly points out that previously, it wasn't said all were serious threats, so why now,does the glass wing seem to add more review workload without much actual benefit?

But in reality, Anthropic did uncover a serious vulnerability in curl.

In September this year, Anthropic used Claude to discover a Use-After-Free memory error in curl. A malicious server could potentially exploit this to achieve remote code execution under certain conditions.

Daniel stated that this was one of the most severe vulnerabilities curl has had in years.

So in the end, VulnCheck's report actually resembles an advertisement for itself, as the issues it criticizes with the glass wing align perfectly with the vulnerability intelligence and priority management services it provides...

One More Thing

At this point, someone might wonder: if Claude is so powerful, why doesn't it just fix the issues directly when it finds them?

In fact, Anthropic has already started generating candidate patches with AI.

However, developers have long been frustrated with AI fixing bugs... "It's not difficult to let LLMs generate fixes, the difficulty lies in whether open-source projects are willing to adopt them."

The workload of reviewing an AI-generated patch can sometimes be almost equivalent to rewriting the code yourself.

often is "lighting a candle to illuminate more darkness", AI discovers a problem and starts fixing it, to fix it, adds more complex logic into the code, finally fixes one bug, but brings ten new problems.

"AI is very good at adding things, but it doesn't realize that things can also be removed..."

At least now, the pits dug by these strongest models still need humans to fill.

Reference link:

https://www.vulncheck.com/blog/anthropic-glasswing-november-update