Free Anthropic OSS Scanner flags bugs with no human review

OSS Scanner is an opt-in service: projects that sign up get thorough, periodic security audits run by Anthropic’s AI at no cost. An AI agent builds the project in an isolated container with no internet access, looks for vulnerabilities and sends a report to the maintainers.

Each report includes a self-contained reproducer, a small test that triggers the bug. Where the model can manage it, the report also includes a suggested patch and a bisection of the code history showing when the flaw first appeared. An early tester from the OpenSSL Corporation said a report with a working exploit attached is basically job done for an engineer.

Anthropic says it was inspired by OSS-Fuzz, the automated bug-hunting service Google and the Open Source Security Foundation (OpenSSF) have run since 2016. The scanner is paid for by Anthropic’s Defender Advantage Fund. The company also sells a paid product, Claude Security, that scans and patches code for businesses.

Every report is fully generated by the model. Anthropic says skipping human triage means faster and more frequent scans, but also that some reports could be incorrect or invalid, for example with the wrong severity rating.

How often does that happen? Anthropic had its own expert penetration testers go through 97 serious findings, rated high or critical, that an early build produced across 48 projects. They approved 85 for disclosure. Of the 12 they held back, 11 were genuine bugs that simply repeated a known issue or another finding. Encryption library maker wolfSSL found only two bad reports among the 74 it got while testing the scanner, and five turned into official CVEs (entries in the public list of known vulnerabilities).

The scanner also exists because of a backlog. Of the more than 29,000 candidate vulnerabilities Anthropic’s models found over six months, staff had manually reviewed about 6,000. Many maintainers who got early reports asked for the unreviewed batch too, and nearly 5,000 such reports have already gone out. Those efforts produced 584 security advisories as of October 2, 2026.

Core maintainers enroll by opening a pull request on the OSS Scanner GitHub repository with a YAML configuration file. It must include:

• a link to the git repository to clone;
• a primary contact email address;
• the path to a Dockerfile that installs all dependencies and builds the project, so the agent can work offline.

Optional extras include more email addresses to copy on reports, the project home page, a GPG public key to encrypt report emails, and a threat model file describing what to test and how to classify issues. Projects can also switch off reports with a single setting. Anthropic recommends checking that the project’s tests pass inside the built container.

Projects are picked using criteria similar to OSS-Fuzz, which looks for critical impact on infrastructure and user security, and decisions are made case by case. Maintainers may need to add a short note explaining why their project matters if it isn’t obvious. By early coverage of the launch, 116 pull requests had already been submitted. Projects without the people to handle raw AI findings can still get human-verified reports through Anthropic’s existing disclosure program.

Bug-reporting programs usually give developers 90 days to fix a flaw before it’s made public. Anthropic says it won’t impose that deadline on OSS Scanner findings, because some may be false alarms. If its team later confirms a report by hand, it may disclose it 90 days after telling the maintainers a human has validated it. The company says it may add disclosure deadlines for some high-severity reports as it gains confidence in the scanner.

If it works as intended, the apps and services you rely on get patched before attackers find the same holes. AI has made finding bugs much cheaper, for defenders and for criminals. Anthropic says it expects AI to favor defenders within two years, making it easier to catch bugs before software ships.

The risk is overload. Some open-source maintainers are already struggling with a wave of AI-generated bug reports, and unreviewed findings add to their workload. Anthropic launched the scanner alongside a Critical Infrastructure Defense Program with 11 founding partners, including CrowdStrike, Palo Alto Networks and Deloitte, aimed at power grids and water systems. Whether the free scans help more than they burden volunteer maintainers is the question to watch over the next few months.