Anthropic's Free OSS Scanner Has a Trust Problem

Anthropic launched a free AI vulnerability scanner for open-source projects on Thursday. Powered by Claude Mythos and offered at no cost, OSS Scanner is the first credible free-tier comp to paid code-security products — and the first AI-driven scanner that ships at scale without human review of the reports it produces.

The launch post is direct about the trade-off. The reports are "fully model-generated and do not require human review or triage." The same paragraph notes that the reports "may contain false positives." In the first six months of the program, Anthropic says it has identified 29,000 candidate vulnerabilities across major open-source projects, reported 6,000 of them to maintainers, and produced 584 public advisories as of October 2.

What Anthropic Shipped

OSS Scanner is opt-in. A maintainer enrolls by submitting a pull request to a registry that adds a YAML configuration file. Once enrolled, the project is scanned on a recurring basis. Findings are generated by Claude Mythos and emailed to the maintainer. There is no UI, no dashboard, no audit trail — just an email with a report.

That last part is the design choice the launch post keeps coming back to. Speed over validation. Anthropic frames it as part of a two-year forecast: "Our forecast is that in two years, AI will favor defense." The argument is that the cost of exploiting vulnerabilities has dropped, and the only way to keep up is to make finding them cheap and fast — even if that means accepting false positives in the pipeline.

The 29,000-to-584 ratio is the number to watch. If 98% of what the scanner finds is noise, the model is doing its job. If 50% is real, that is a maintainer-facing problem. Either way, the burden of proof runs through the maintainer, not the model — and the program has no built-in validation step before reports go out the door.

The Trust Stack That Is Missing

Three things are notably absent from the program as designed.

No human review. The reports are "fully model-generated." When Claude Mythos finds a bug, the report ships. When it does not, the report still ships. The maintainer is the one who has to figure out which is which.

No 90-day disclosure clock on AI findings. Standard coordinated vulnerability disclosure runs on a 90-day timer: report, wait 90 days, publish. Anthropic's program explicitly waives this for unvalidated AI reports. The reasoning is reasonable — false positives should not be auto-disclosed — but the side effect is that the company has set no upper bound on how long an AI-generated finding can sit in a maintainer's inbox, or how a hallucinated CVE could leak sideways through GitHub issues, social media, or downstream users before anyone validates it.

The 90-day rule is asymmetric. Anthropic says it will impose a 90-day disclosure period only on human-validated findings. AI-only findings get no clock at all. The asymmetry rewards fast AI shipping and punishes thoroughness, because the human-validated path is the one that gets public-disclosure protection. If you are a maintainer, the cheapest way to keep your name out of a leaked report is to be the kind of project that gets human validation; if you are a vendor with an internal codebase, you get neither.

What This Means for Paid Code Security

When "free" enters a market, the pricing curve compresses. Every paid code-security product is going to be asked the same buyer question this quarter: why pay for this when Anthropic runs the equivalent for free? The honest answer differs product by product. The durable framing is: free AI security tools are not free. The cost is in the trust assumptions baked into the output.

A human-triaged report with a 90-day disclosure clock is a different product from an AI-only report with no clock. The market is going to have to learn the difference, one false-positive incident report at a time. The vendors who already bundle human review, audit trails, private-repo coverage, or supply-chain provenance now have a clear differentiator — not because their product is more capable than Claude Mythos, but because their product makes a different trust promise to the maintainer on the other end.

What Builders Should Take From It

• If you maintain an open-source project: opt in only if you have time to triage what comes in. The default assumption is that the report needs validation before you act on it.
• If you buy code-security tools: the next vendor conversation should be about the trust layer, not the model. Ask who validates findings, what the disclosure clock is, and what the audit trail looks like.
• If you build in this space: the bar has moved. The product is not the scanner; the product is the report. Free is the new feature, but trust is the new moat.
• If you are watching the AI-trust story: this is the canonical case. A frontier model shipped at scale, with the trade-off in plain English, and the maintainer community as the implicit beta tester.

---

Developer312 uses AI tooling, including Anthropic's Claude.