Claude uncovers 38,000 bugs—does it only find them, not fix them?

Claude uncovers 38,000 vulnerabilities, but does it fix them?

This side, AI strikes and immediately digs up everything, first uncovering a major PHP core security vulnerability, then exposing 948 Linux kernel vulnerabilities…

On the other side, the open‑source community is on the brink of collapse: you find the vulnerabilities, now go fix them!

Yesterday, cybersecurity firm VulnCheck released a report on AnthropicGlasswing Project (Project Glasswing).

The data shows that seven months after the launch of the Glasswing Project, Claude has identified a total of37,966 candidate vulnerabilities.

However, to date, only 217 have been publicly recorded as fixed.

VulnCheck can’t help but wonder: did the much‑hyped cybersecurity “Mythos era” ever truly arrive?

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

First, let’s remind everyone of Anthropic’s high‑profile “Glass Wing” program launched in April this year.

This initiative aims to leverage Claude to massively uncover security vulnerabilities in critical software and open‑source projects, collaborate with tech companies and open‑source maintainers to drive fixes, and eliminate threats before they are exploited.

The name carries two meanings.

On one hand, security bugs are like glass‑winged butterflies, hidden in plain sight for decades without detection; on the other hand, transparency itself is a form of protection. Anthropic hopes that more transparent security research and vulnerability disclosure mechanisms will give software systems better defense.

△Glasswing butterfly, scientific name: Broad-striped black-veined fritillary | Image source: Wikipedia

At launch, Anthropic claimed it would provide $100 million in model usage credits and $4 million in open-source security donations, and also announced its flagship cybersecurity model Claude Mythos Preview has discovered thousands of high‑risk and critical vulnerabilities. Subsequently, Apple, Google, Microsoft, AWS, and NVIDIA all joined.

But is it really so?

According to the continuously tracked data from VulnCheck, as of early November, Anthropic has accumulated 37,966 candidate findings, but only 217 of them are marked as fixed in the vulnerability disclosure ledger.

Meanwhile, obsolete data continues to pile up, with 419 vulnerability retractions and as many as 1,654 duplicate entries.

△Image source: VulnCheck

This gap is still widening rapidly.

In just the past two months, candidate vulnerabilities have surged by nearly 12,000, but the number of fixed records has only barely increased by 15.

Additionally, the 442 CVEs attributed to Anthropic tracked by VulnCheck account for only 0.5% of the total global public CVEs in 2026.

△Image source: VulnCheck

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

Although the number of fixes varies slightly due to different data scopes, there is one thing both sides have discovered.

That is, the intensity of AI-driven vulnerability hunting is simply overwhelming for manual review.

Anthropic has long recognized this, and the company admits that the severe bottleneck in 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, allowing the model‑generated vulnerability reports to be sent directly to the maintainers without waiting for Anthropic’s manual review.

Open‑source maintainers can’t keep up

AI is forcing us to pay off 40 years of technical debt right now.

Behind this poor data, the ones who suffer most are open‑source maintainers, such as VulnCheck’s example of the network data transfer tool curl.

In early November, curl maintainer Daniel Stenberg received another big gift from Anthropic, 45 vulnerability reports.

And even before that, Daniel wrote in his blog that core maintainers are being overwhelmed by a constant flood of reports.

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

To handle so many reports, Daniel announced a change to curl’s security vulnerability handling policy. Low- and medium‑severity vulnerabilities will now be discussed publicly on GitHub,so that more community developers can help review.

Currently, out of the 34 publicly disclosed curl reports, 19 have been determined not to constitute security vulnerabilities, and 6 have been confirmed as low‑severity issues.

VulnCheck’s report subtly notes, “Earlier we said they were all severe threats, but now,” “GlassWing seems to add extra review workload without much benefit?”

In fact, Anthropic did uncover a serious vulnerability in curl.

In September this year, Anthropic, using Claude, discovered a Use‑After‑Free memory error in curl. Under certain conditions, a malicious server could even achieve remote code execution.

Daniel said it is one of the most severe curl vulnerabilities in years.

So, back to the point, VulnCheck’s report is basically an ad for themselves, because the issues it criticizes in GlassWing happen to align closely with the vulnerability intelligence and prioritization services they offer…

One More Thing

Having said that, some might wonder: since Claude is so powerful, why not just fix the issue directly when it’s discovered?

In fact, Anthropic has already gotten AI to generate candidate patches.

However, developers have been struggling with AI bug fixing for a long time… “Getting an LLM to generate a fix is indeed easy; the hard part is whether open‑source projects are willing to adopt it.”

Reviewing an AI‑generated patch can sometimes take almost as much effort as writing one from scratch.

Often it’s “light a candle to illuminate more darkness”; AI finds a problem and starts fixing it, and to fix it adds more complex logic to the code, finally fixing one bug but bringing ten new issues.

“AI is very good at adding things, but it doesn’t realize that things can also be removed…”

For now, the pitfalls these most powerful models dig still need humans to fill.

Reference link:

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