Anthropic will scan critical open-source projects for free, and the reports arrive without a human in the loop

Anthropic is offering free, recurring vulnerability scans to critical open-source projects, and it has removed the step that usually slows that kind of work down. OSS Scanner, announced October 8, emails model-generated findings straight to maintainers without a human reviewing them first. The bet appears to be that faster reports are worth some wrong ones, and the cost of sorting the two now lands on the projects that opt in.

What Anthropic shipped

OSS Scanner is part of a broader announcement Anthropic calls its Cyber Mission, which also includes a defense program for critical-infrastructure security vendors. For developers, the scanner is the piece with a direct effect.

Anthropic says enrolled projects get “periodic scans from our most capable models, free of charge.” Each report includes a proof of concept showing how the bug could be exploited, an explanation, and a suggested fix where one is available. Anthropic says it expects “a true-positive rate above 90%,” and it is explicit about the trade-off: “The reports are model-generated and sent without human review. That means maintainers receive them faster, but it also means that some will contain inaccuracies.”

Eligibility borrows the test from Google’s OSS-Fuzz: established projects with a critical impact on infrastructure and user security. Enrollment is a pull request to the anthropics/oss-scanner GitHub repository that adds a project.yaml, a Dockerfile that builds the project, and an optional threat model. The scanner builds the project in an isolated VM, then analyzes it with no internet access. Anthropic says it will not apply a 90-day disclosure deadline to these unreviewed findings and will not make them public.

The numbers behind the bet

The case for skipping human review rests on early accuracy data that, for now, comes mostly from Anthropic. Help Net Security and SiliconANGLE both report that penetration testers assessed 97 high and critical findings across 48 projects from an early version of the scanner, with 85 meeting Anthropic’s disclosure bar, 11 real but duplicates of known issues, and one invalid. SiliconANGLE also reports that wolfSSL found 72 of 74 early reports valid, with five becoming CVEs, and Help Net Security quotes an OpenSSL Corporation engineer saying the AI-generated reports matched or beat human-written ones, especially when an exploit came attached.

Those are encouraging signals from a small, early sample. They do not yet show how the true-positive rate holds up across hundreds of projects with different languages, build systems, and threat models.

What it means for maintainers

For an enrolled project, the practical change is a new inbound queue of security reports that look credible and may or may not be. The optional threat_model.md matters more than it first appears. Anthropic says it is most useful for telling the scanner how the project rates severity, where untrusted input enters, and what is out of scope. A project that skips it is likely to receive more reports it considers noise.

Maintainers should weigh capacity before enrolling. Anthropic itself frames the service as meant for “projects with the capacity to keep up with surfaced findings.” A project already struggling to answer human reports will not be helped by a faster stream of machine ones, even accurate ones.

The reproducer is the useful part. A report that ships with a working proof of concept can be confirmed or dismissed far faster than a prose-only report, which is what makes an unreviewed stream workable at all.

What it means for everyone downstream

Most developers will never enroll a project, but many depend on the libraries that will. If the scanner works as described, the near-term effect is more security releases from core dependencies such as TLS libraries, parsers, and language runtimes. Teams that pin versions loosely or patch on a slow cycle should expect that cadence to pick up, and dependency update tooling becomes more important rather than less.

The industry question is less whether AI can find bugs and more whether open-source maintainers can absorb what it finds.

What to watch

The signals worth tracking are how many projects enroll, whether maintainers publicly report true-positive rates near Anthropic’s 90 percent expectation, how many CVEs cite the scanner’s ANT- identifiers, and whether other model vendors follow with similar programs. Stackmaven’s follow-up will revisit around January 9.

Sources cited

1. Introducing the Anthropic Cyber Mission (Anthropic) www.anthropic.com
2. OSS Scanner (Anthropic) red.anthropic.com
3. anthropics/oss-scanner enrollment repository (GitHub) github.com
4. Anthropic offers free AI security scans to open-source maintainers (Help Net Security) www.helpnetsecurity.com
5. Anthropic launches critical infrastructure program and free OSS Scanner for open source (SiliconANGLE) siliconangle.com