🔥 Trending on HN

Anthropic’s AI Found Weaknesses. Who Can Check Them?

3 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Project Glasswing(Project Glasswing)

Anthropic’s project for finding software weaknesses with AI.

OSS Scanner(OSS Scanner)

An opt-in service that sends AI-generated security reports to open-source projects.

true-positive rate(true-positive rate)

The share of reports judged real under a stated method.

What happened

Anthropic, the company that makes Claude, began Project Glasswing in April 2026. It used AI to search open-source software for security weaknesses. A November 9 review by VulnCheck, a vulnerability-intelligence company, tracked 442 CVE entries attributed to Anthropic over seven months. That number does not mean 442 new, confirmed, high-impact flaws. It is VulnCheck’s own tracking set. Read the VulnCheck review

The story also drew attention on Hacker News. The snapshot used for this article showed 359 points and 164 comments. Those numbers measure community attention. They do not prove that the source is correct or that the flaws are dangerous. See the Hacker News thread

The background

Anthropic’s October 8 announcement described a growing validation problem. The company said its systems had found more than 29,000 candidate vulnerabilities in six months. People had manually reviewed about 6,000. Anthropic also said it had sent nearly 5,000 unverified reports to maintainers who requested them.

Anthropic then launched OSS Scanner. Projects can opt in to receive recurring scans. The reports are generated by models without human review or triage. Anthropic warns that some reports may be wrong or invalid. Anthropic’s announcement

What the available numbers show

Anthropic’s ledger, updated November 2, listed 37,966 findings. It showed 7,349 as public and 217 as fixed. About 1,600 were marked as duplicates, and 419 were withdrawn. The ledger reported a 94.2% true-positive rate. These are ledger categories, not a count of 37,966 confirmed vulnerabilities.

VulnCheck’s tracker found that two of its 442 CVEs had been exploited in the wild. The review also examined a curl sample. curl is a widely used open-source software tool. Anthropic sent curl 45 reports. Of 34 published reports, 19 were marked not vulnerabilities, six were confirmed as low severity, and nine awaited assessment. This is an interim sample from one project. It cannot establish the program’s overall accuracy.

Anthropic presents a different test. It says expert testers accepted 85 of 97 high- or critical-severity scanner findings for its disclosure process. That sample used different selection rules, so it is not a direct comparison with curl.

Why this matters

The central issue is not only whether AI can find flaws. Someone must decide whether each report is real, judge its impact, prepare a fix, and coordinate disclosure. If discovery grows faster than validation, the tool can create a large queue for the people it aims to help.

VulnCheck argues that Glasswing’s burden has shifted toward open-source maintainers. That is the author’s assessment, not a proven result for every project. The reports are not useless, either. curl’s maintainer previously confirmed one low-severity vulnerability from an Anthropic Mythos scan. curl maintainer’s account Useful findings and costly noise can arrive together.

What remains uncertain

The sources do not fully explain how Anthropic defines a finding, a true positive, a duplicate, or a fixed report. VulnCheck also says the ledger may have a display bug that inflates disclosed results. That claim needs independent confirmation. We also do not know how many pending reports will become fixes, withdrawals, or duplicates.

What to watch next

Future ledger updates should show whether fixes rise with public reports. Independent maintainer checks matter more than raw discovery totals. Watch whether OSS Scanner adds human review, better severity estimates, or support for the teams doing the triage. The real test is not how many weaknesses AI can name. It is how many people can safely verify and repair.

💬 What Glasswing’s numbers do—and do not—show

The central debate is that Glasswing may show strong vulnerability discovery, but its figures alone do not prove that the findings are dangerous, exploited, or successfully fixed.

  • As commenters relayed the article’s figures, Anthropic reported 37,966 findings; 7,349 were public (19.4%), and 217 were marked fixed (about 0.5% of the total).
  • One commenter notes that this is analysis of only the roughly 20% that became public. The other 80% could include unresolved findings, non-issues, duplicates, or withdrawn reports, so the public sample may be heavily selected.
  • In self-reported experience from an OSS-scanner user, more than 20–30% of the issues were junk and many suggested patches were unusable. Another user’s self-report put false positives near 50%, with about 20% of the remainder being bugs rather than vulnerabilities; only a small fraction of patches were accepted.
  • This supports the concern that AI discovery may be scaling faster than reliable remediation, shifting review work onto maintainers. Detection volume is not enough; human review time and patch quality matter too.
  • On PHP2Shell, one side argues that difficulty for an ordinary user does not disprove serious risk: capable attackers may have established methods for exploiting hard vulnerabilities.
  • A commenter reported that an AI agent took digital assets from a deliberately vulnerable test site. The initial report said two hours, but another commenter corrected the timing to 19.5 hours. This is an anecdotal test report, not evidence of widespread real-world exploitation.
  • The opposing view is that the predicted mass incident—thousands of sites being compromised within hours—did not occur, so vulnerability existence and successful exploitation must be separated.
  • There is no consensus. The useful comparison is among public-data bias, false positives, patch quality, reproducible exploits, and confirmed impact—not just the number of findings.

initial digest at 216 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

AI Finds Weak Spots. People Still Need to Check Them.

📰 Full story: Anthropic’s AI Found Weaknesses. Who Can Check Them?

AI can find software weaknesses quickly. People must still check and fix the reports.

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
CVE(CVE)

A public label used to identify a software weakness.

OSS Scanner(OSS Scanner)

A service that sends AI-made reports to projects that choose it.

maintainer(maintainer)

A person or team that looks after software.

💡 The gist

  • Anthropic uses AI to find weak spots in shared software.
  • Many reports still need human checking before anyone fixes them.
  • Hacker News discussed the story widely; attention is not proof.

Anthropic makes Claude. It started Project Glasswing to scan open-source software. Open-source software can be shared, changed, and improved by many people.

VulnCheck tracked 442 CVEs linked to Anthropic. A CVE is a public ID for a software weakness. Its review found two used in real attacks. This does not mean the other 440 were useless. It means the review confirmed attacks for only two.

Anthropic’s ledger listed 37,966 findings on November 2. It listed 7,349 as public. It listed 217 as fixed. About 1,600 were duplicates. Another 419 were withdrawn. A finding is not automatically a confirmed weakness.

Anthropic also launched OSS Scanner. Projects can choose to receive its reports. The AI writes these reports without a person checking first. This makes scanning faster. It can also send more work to maintainers.

curl, an open-source tool, received 45 reports from Anthropic. Of 34 published reports, 19 were not vulnerabilities. Six were low severity. Nine were still being checked. This was only one project’s early sample. It cannot measure every Anthropic report.

Anthropic reports a different test. Experts accepted 85 of 97 serious findings. The test used a different group of reports. So the two results cannot be compared directly.

The key question is simple. Can people check reports as quickly as AI creates them? Checking means testing the report, judging its danger, and making a repair. If reports arrive too quickly, the queue grows.

VulnCheck says this work is moving to open-source maintainers. That is its judgment. It is not proven for every project. AI reports can still help. curl previously confirmed one low-severity flaw from an Anthropic Mythos scan. curl’s account

The next evidence will come from new ledger updates. Watch fixed reports, duplicate reports, and independent checks. Also watch whether OSS Scanner adds more human support.

In the supplied Hacker News snapshot, the post had 359 points and 164 comments. That shows interest. It does not show that the claims are true. VulnCheck review Anthropic announcement Hacker News thread

💬 Are AI-found vulnerabilities really dangerous?

The comments argue that finding many weaknesses does not automatically mean they are dangerous, exploitable, or properly fixed.

  • The article’s figures, as repeated in the thread, say there were 37,966 findings, 7,349 public reports, and 217 marked fixed. The hidden 80% is unknown.
  • In users’ self-reported experience, many findings were wrong or low-value. One report said 20–30% or more were junk; another said about half were false positives and only a small share of patches were usable.
  • So AI may be better at spotting possible problems than producing trusted fixes. Maintainers still have to spend time checking both the report and the patch.
  • PHP2Shell is disputed. Some commenters say a hard-to-use vulnerability can still be serious, while others point out that no huge wave of attacks appeared. A test-site report was also corrected from two hours to 19.5 hours.
  • The key distinction is between a vulnerability, a working exploit, and confirmed widespread harm. Finding counts alone cannot settle the question.

initial digest at 216 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

AI Found Weak Spots. Who Checks Them?

📰 Full story: Anthropic’s AI Found Weaknesses. Who Can Check Them?

AI can find many weak spots. People must check them before repairs.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Anthropic(Anthropic)

The company that makes Claude.

CVE(CVE)

A name used for a software weak spot.

Hacker News(Hacker News)

A website where people discuss technology.

A simple story

Anthropic makes Claude, an AI helper. It asked AI to look for weak spots in software. A weak spot is a place that can cause trouble.

The AI found many possible problems. People still need to check each one. They decide if it is real. Then they choose how to fix it. Some reports arrive before a person checks them.

Anthropic’s list had 37,966 findings. It marked 217 as fixed. About 1,600 were repeated findings. It removed 419 findings. So, not every finding was a real problem.

VulnCheck tracked 442 linked CVEs. It found two used in attacks. A CVE is a name for a software weakness.

curl, a software tool, received 45 reports. People checked 34 of them. Nineteen were not weaknesses. Six were small weaknesses. Nine still needed checking.

Hacker News is a place for tech discussions. Its snapshot showed 359 points and 164 comments. That means many people noticed the story. It does not make the story true.

AI can search very quickly. People cannot check everything instantly. The next question is simple. Will fixed problems grow faster than new reports? The VulnCheck review The Hacker News thread

💬 AI found many weak spots, but that is not the whole story

A big number of computer weaknesses does not tell us how many are truly dangerous.

  • The article’s figures say about 20% of the reports were public and 217 were marked fixed out of 37,966. We do not know what most of the hidden reports contain.
  • Users reported that many findings were mistakes and that the suggested fixes often needed careful human checking.
  • People disagree about PHP2Shell: some say a hard weakness can still hurt people, while others say there was no giant attack. A test-site report was corrected from two hours to 19.5 hours.
  • A weakness, a real attack, and a widespread disaster are different things. We need to check all three.

initial digest at 216 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

Sources