Anthropic Says Claude Helped Surface 2,300 Disclosed Open-Source Vulnerabilities
Anthropic’s coordinated disclosure dashboard says Claude models helped identify vulnerabilities that led to 2,300 disclosures across 392 open-source projects, with human triage and maintainer review built into the process.
Anthropic publishes a large-scale vulnerability disclosure ledger
Anthropic has published an updated coordinated vulnerability disclosure dashboard describing a security research program that uses Claude models to find candidate vulnerabilities in open-source software, then routes those findings through human triage, external security review, and maintainer disclosure.
The dashboard was last updated on August 26, 2026. Anthropic says it began the effort in February 2026 with an early snapshot of Claude Mythos Preview, later including other Mythos-class and Claude models.
As of the August 26 update, Anthropic reports that the program has disclosed 2,300 vulnerabilities across 392 open-source projects. The company says 421 have been patched upstream, while 462 identifiers have been issued across the ledger: 177 CVEs and 285 GitHub Security Advisories. A single finding can carry both identifiers.
This is a security-research and coordinated-disclosure program, not a claim that Claude autonomously validates or fixes vulnerabilities without human review.
How the disclosure pipeline works
Anthropic’s process starts with model-generated candidate findings. The dashboard currently shows 26,153 discovered findings entering the broader pipeline and 5,008 candidate findings moving into deeper review.
For most disclosures, Anthropic or one of six external security research firms attempts to reproduce each issue, determines whether it is a real bug, assesses severity, and prepares a report for the project maintainer. Anthropic says 4,576 findings have been reviewed by external security firms.
The dashboard reports a 91.4% true-positive rate for the manually reviewed candidate set represented in the flow. Anthropic cautions that this figure is only one proxy for impact: a technically valid bug may already be known, fall outside a project’s threat model, or ultimately receive no fix. The number of upstream patches is therefore a more concrete but slower-moving signal.
Some maintainers ask Anthropic to disclose untriaged findings directly. Those reports do not receive the same independent external review before disclosure, and the dashboard labels that path separately.
Why the public ledger matters
The dashboard does more than publish aggregate counts. Anthropic maintains a disclosure ledger with a SHA-3-512 hash commitment for validated findings that are still inside the coordinated disclosure window.
Before a finding can safely be revealed, the ledger can publish its commitment hash and date without exposing exploit details. As the disclosure process advances, status and severity information are added. Project names, identifiers, and bug classes are revealed after the disclosure window closes.
That structure is important because security teams need a way to demonstrate when a finding existed without publishing sensitive technical details before maintainers have had time to patch.
The public records already include vulnerabilities affecting widely used projects and infrastructure components. However, readers should rely on the underlying CVE, GHSA, maintainer advisory, and patched release for operational remediation rather than treating the dashboard itself as a substitute for vendor security guidance.
What the numbers do and do not show
The scale of the program is notable, but the numbers require careful interpretation.
Anthropic says model discovery is not the rate-limiting step. Independent human triage and review are the bottleneck, which means the 2,300 disclosed vulnerabilities are only a subset of all candidate bugs the models have produced.
A model-generated finding is also not automatically equivalent to a confirmed vulnerability. Reproduction, severity assessment, deduplication, maintainer response, and patch development remain human- and project-dependent.
The dashboard’s patch count likewise does not mean every downstream user has installed a secure version. “Patched upstream” means the maintainer has created and released a fix; organizations still need to identify affected software and apply the relevant update.
Why this matters for AI-assisted security
The program provides one of the clearest public examples yet of frontier models being used at scale for defensive vulnerability research rather than only benchmark demonstrations.
Its most important lesson may be operational rather than purely model-centric: AI can generate enough plausible findings that the surrounding verification and disclosure system becomes the limiting factor. That shifts attention toward triage capacity, reproducibility, secure evidence handling, maintainer coordination, and responsible disclosure governance.
For security engineering teams, the dashboard is also a useful template for separating three stages that are often blurred together in AI-security claims:
- Discovery — a model proposes a potentially exploitable condition.
- Validation — humans reproduce the issue and assess whether it is real.
- Remediation — maintainers acknowledge, fix, disclose, and distribute the patch.
Anthropic’s August update shows substantial throughput across all three stages, while also making the boundaries visible.
Release status
The dashboard is live and publicly accessible. It is an ongoing disclosure tracker, not a standalone commercial security product announcement. Anthropic says the underlying model-assisted research continues, and the ledger will change as more findings are reviewed, disclosed, assigned identifiers, and patched.
This article is built from the source material below. Open the originals for full context and the latest updates.
