Anthropic has opened a new front in the fight against software vulnerabilities, and it is putting its strongest AI models on the case. The company launched OSS Scanner on October 8 as a free, opt-in service for critical open-source projects, sending findings to maintainers without human review.
That design creates a striking trade: projects can receive security findings faster, but their teams must handle triage and validation themselves. Just 11 days after launch, Anthropic reported 77 critical and 803 high model-rated findings across the program.
A Free Security Audit Built for Open Source
OSS Scanner runs on Anthropic’s infrastructure and uses the company’s strongest models, including Claude Mythos, to audit enrolled repositories. The Frontier Red Team and Safeguards Research team run the program together, with a design that borrows from Google’s OSS-Fuzz but swaps traditional fuzzers for language models.
Enrollment is open to core maintainers through a pull request on anthropics/oss-scanner. The pull request adds projects/<project>/project.yaml, based on the project template, and requires three fields:
- repo
- primary_contact
- A Dockerfile path
Maintainers can also provide a threat model, a PGP key for encrypted reports, automatic carbon copies, a homepage, and a disabled setting. An optional threat_model.md can define adversarial inputs, code that falls outside the audit, and a severity rubric. Without that file, the scanner guesses.
The Dockerfile builds the project with network access, but the audit itself takes place inside an isolated virtual machine with no Internet. Agents in the pipeline double-check bugs, perform root cause analysis, and propose patches before delivering each report.
Each report includes a self-contained reproducer and an explanation of the bug. Where possible, it also identifies the commit that introduced the issue through a bisection, and it attaches a candidate patch when one is available.
Early Results Show Strong Detection, With Important Limits
As of October 19, OSS Scanner had accepted 114 projects from about 650 applications. The program’s early scale is small compared with the number of projects that depend on open-source code, but its first performance review produced a clear signal.
Anthropic asked its penetration testers to examine 97 critical and high findings across 48 projects. The testers judged 85 of them, or 88%, strong enough to meet Anthropic’s bar for its coordinated vulnerability disclosure process. Only 1 of the 97 findings was a false positive.
The remaining results show why maintainers still need to lead the response. Eleven findings were real but duplicated known issues or other scan findings, and maintainers reported inflated severity in those cases. A high model-rated result can point to a genuine problem while still requiring human judgment about its scope, impact, and priority.
Maintainer feedback points in the same direction. wolfSSL received 74 reports, and 72 were valid. Five became CVEs, turning model-generated alerts into formal vulnerability records that project teams could address.
curl also patched CVE-2026-92392, a High-severity HTTP/2 upload use-after-free. The issue affects curl 8.12.0 to 8.22.0 and is fixed in curl 8.23.0, released October 14, 2026.
From Experimental Scans to a Standing Security Pipeline
OSS Scanner turns a process that had grown through individual requests into an opt-in program. Over 6 months, Anthropic’s models surfaced over 29,000 candidate vulnerabilities. Human reviewers triaged roughly 6,000 of them, while nearly 5,000 more reports went straight to maintainers who asked to receive everything, whether validated or not.
The new system gives that demand a formal path. Maintainers choose to enroll, define the project build, and decide whether to provide extra information that can guide the models. Anthropic then runs periodic audits and delivers the results without human review before delivery.
Anthropic links the scanner’s results to progress on CyberGym, where language models went from finding under 20% of vulnerabilities to finding over 85%. The company’s broader coordinated vulnerability disclosure dashboard listed 6,157 disclosed vulnerabilities across 591 projects as of October 2, with 516 of those patched.
Those numbers frame OSS Scanner as more than a one-time experiment. The program combines model-powered discovery, isolated execution, reproducible reports, root cause analysis, and proposed fixes, but it leaves the final security decision with project maintainers.
That balance will shape the next stage of AI-assisted security. More enrolled projects could produce more findings, and more findings could bring more pressure to separate urgent flaws from duplicates, severity errors, and issues that need additional validation. For open-source teams facing limited time and large codebases, a free stream of detailed reports could become a powerful advantage—provided they have the capacity to act on what the scanner finds.
Based on
- Anthropic Launches OSS Scanner, a Free AI Vulnerability Scanner — marktechpost.com
- Anthropic’s Mythos Kept Hacking Even When It Knew People Could Get Hurt – Business Insider — businessinsider.com
- Evaluators say ‘vulnerabilities’ found with OpenAI monitoring system – UPI.com — upi.com
- AI Is Exposing the Internet’s Human Slop Problem – Business Insider — businessinsider.com
- What smart people are saying about the AI cybersecurity ‘vulnerability tsunami’ — africa.businessinsider.com
