# Anthropic's Claude Found 29,000 Open-Source Vulnerability Candidates—Why Only 516 Are Fixed

Anthropic has launched OSS Scanner, a free automated vulnerability scanning service for critical open-source software powered by its top frontier models including Claude Mythos, after internal telemetry exposed a massive bottleneck in global software security: Claude is discovering real security bugs far faster than human researchers can validate them and open-source maintainers can patch them. Over six months of scanning foundational open-source repositories, Anthropic's models surfaced more than 29,000 candidate vulnerabilities, yet only 516 have been patched upstream so far.

Digging into the numbers on Anthropic's new vulnerability disclosure dashboard reveals that this is not another case of AI spamming maintainers with hallucinated false positives. To vet the 29,000 candidate flaws, Anthropic contracted six external security research firms to manually review the model's output. As of October 2, 2026, those independent human researchers had worked through 6,123 of the 29,000 candidates—and confirmed that 5,674 of them (an extraordinary 92.7% validation rate) were genuine vulnerabilities, while roughly 23,000 candidates remain sitting in the unreviewed queue.

Even more striking is what happened after Anthropic sent 6,157 validated findings (resulting in 584 CVE and GitHub Security Advisory identifiers) to upstream maintainers: only 516 vulnerabilities have been fixed in production releases, leaving more than 90% of confirmed bugs unpatched.

Why is the open-source ecosystem struggling to absorb 5,674 real bug reports, and why did cURL founder Daniel Stenberg—who famously shut down cURL's bug bounty program in January 2026 due to a flood of low-quality AI-generated slop—publicly endorse Anthropic's OSS Scanner? Unlike generic static-analysis (SAST) alerts, Anthropic runs each target project inside an isolated, internet-disconnected virtual machine and packages every report with three high-signal artifacts: a self-contained exploit reproducer, a `git bisect` analysis pinpointing the exact commit that introduced the bug, and a proposed remediation patch.

In early testing across 48 projects, expert penetration testers confirmed that 85 of 97 critical and high-severity scanner findings cleared the bar for coordinated disclosure. WolfSSL founder Todd Ouska reported that 72 of the 74 reports his team received were valid (with five becoming CVEs), and PostgreSQL committer Noah Misch noted that several proposed patches arrived ready to merge "nearly as-is."

As detailed in [The New Stack's report on Anthropic's OSS Scanner and the 29,000-bug backlog](https://thenewstack.io/anthropic-oss-scanner-vulnerabilities/), because human validation cannot keep pace with a 23,000-bug queue, Anthropic now lets verified core maintainers of OSS-Fuzz-eligible projects opt into receiving unvalidated Claude Mythos reports directly—with no 90-day disclosure clock on unvalidated findings—and has already delivered nearly 5,000 raw reports to maintainers who asked for everything the scanner found.

## Fast Facts: Anthropic OSS Scanner Telemetry (Oct 2026)

- **Total AI-Surfaced Candidates**: 29,000+ potential vulnerabilities discovered across critical open-source projects over six months.
- **Human Validation Accuracy**: Outside security firms reviewed 6,123 candidates and confirmed 5,674 as valid bugs (92.7% true-positive rate); ~23,000 remain unreviewed.
- **Critical/High-Severity Precision**: On an early sample of 97 critical and high-severity findings across 48 projects, 85 passed expert pen-tester vetting (87.6%).
- **Upstream Patch Bottleneck**: Out of 6,157 findings sent to maintainers (584 CVEs/GHSAs issued), only 516 have been patched to date (<9% remediation rate).
- **Maintainer Validation**: WolfSSL confirmed 72 of 74 reports were valid (5 CVEs); endorsed by PostgreSQL committer Noah Misch and cURL founder Daniel Stenberg.

## The AI Vulnerability Funnel: Discovery vs. Human Patch Capacity

| Pipeline Stage | Volume / Metric | Conversion / Validation Rate | Primary Operational Bottleneck |
|---|---|---|---|
| **1. Claude Model Discovery** | 29,000+ candidate bugs (6 mos) | Baseline AI discovery volume | ~23,000 candidates stuck waiting for human review |
| **2. External Firm Triage** | 6,123 reviewed -> 5,674 valid | 92.7% true-positive accuracy | 6 external security firms cannot match model speed |
| **3. Maintainer Notification** | 6,157 sent (+ ~5,000 raw opt-in) | 584 CVEs / GHSAs assigned | Volunteer/small maintainer teams overwhelmed by volume |
| **4. Upstream Patch Merged** | 516 bugs fixed in release | 8.4% of reported bugs patched | Regression testing, backporting, and breaking-change risk |

## Why Reproducers and Patches Still Leave a 91% Remediation Gap

If Anthropic's scanner includes a working reproducer, a commit bisection, and a proposed fix that PostgreSQL committers say works "nearly as-is," why have only 516 of 6,157 reported bugs been patched? As OpenSSH maintainers noted during recent disclosures, fixing a deep memory or protocol edge case in foundational C libraries often alters legacy behavior that downstream enterprise applications implicitly rely upon. Maintainers must still review edge cases, run full regression suites, backport patches across three or four supported release branches, and coordinate release windows.

For enterprise CISO teams, the takeaway is stark: there are currently more than 5,100 confirmed, valid vulnerabilities—and roughly 21,000 additional statistically likely bugs inside the 23,000 unreviewed queue—sitting inside the open-source dependencies your production containers run every day.

### The 4-Step Enterprise Implementation Playbook

1. **Enroll Eligible Corporate Open-Source Projects in Anthropic OSS Scanner**: If your engineering organization maintains widely used open-source libraries that meet OSS-Fuzz eligibility criteria, have a core maintainer apply via Anthropic's GitHub repository for free periodic Claude Mythos scans in isolated VMs.
2. **Require Reproducer + Bisect + Patch Packaging from Internal AppSec Tools**: Upgrade your internal security standards to match the OSS Scanner format—ban AppSec scanners from filing raw SAST tickets into Jira unless an agent has first verified the bug in an isolated VM, identified the introducing commit via `git bisect`, and drafted a candidate fix.
3. **Deploy Virtual Patching and Runtime Sandboxing for Unpatched Upstream Dependencies**: Because open-source maintainers have only patched 516 of 6,157 confirmed bugs so far, assume your upstream C/C++ libraries contain unpatched flaws; isolate parsers, cryptographic libraries, and network daemons inside minimal privileges and WAF/eBPF runtime filters.
4. **Sponsor or Dedicate Engineering Hours to Upstream Patch Backporting**: Rather than waiting months for volunteer maintainers to regression-test fixes on libraries your business depends on, assign internal staff engineers to help upstream projects validate and backport security patches.

  Next Steps

1. Inventory your Tier-1 open-source C/C++ and Rust dependencies (such as wolfSSL, PostgreSQL, cURL, and OpenSSH) and subscribe to their security advisory feeds to apply the 584+ newly issued CVEs and GHSAs immediately.
2. Check if any open-source projects maintained by your company qualify for Anthropic's free OSS Scanner program.
3. Configure your internal CI pipeline to run automated `git bisect` and sandboxed reproducer verification on all high-severity vulnerability alerts.