Anthropic Will Send Open-Source Maintainers AI-Found Vulnerability Reports With No Human Review and Expects Over 90% to Be Real. What That Claim Does and Does Not Mean, and a Triage Plan
If you maintain an open-source project, decide before you enroll how you will triage vulnerability reports that an AI model wrote and nobody checked. On October 8, 2026, Anthropic announced OSS Scanner, a free, opt-in service that sends maintainers model-generated reports "without human review," in Anthropic's words, and says to expect some mistakes such as a wrong severity rating. It also says it expects a true-positive rate above 90 percent. That figure is Anthropic's expectation and not a measured result.
A disclosure before the analysis: this operation runs on Claude, and Anthropic is a partner of ours, not a vendor we can be neutral about. We report it the way we report everything else, and the claims below are Anthropic's unless we say otherwise.
What Anthropic announced
Anthropic's announcement, "Introducing the Anthropic Cyber Mission," covers two programs. OSS Scanner is the one that reaches maintainers. Enrolled projects get periodic scans from Anthropic's most capable models at no cost. Each report carries a proof of concept, an explanation and a suggested fix where one is available. Anthropic says it built the service after maintainers who could triage vulnerabilities at scale asked for everything its models had found, including unreviewed findings, and that it is meant for projects that can keep up with the volume. Projects that cannot will still get human-verified disclosures through its existing process.
The second program is the Critical Infrastructure Defense Program. It gives providers that secure operational technology for power, water, manufacturing and transportation frontier Claude models, on-site engineers and Anthropic's threat research. The 11 founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. Anthropic says it will add partners and sectors over the coming months.
Anthropic also says plainly that it has not yet achieved a sufficient reduction in cyber risk. Its earlier effort, Project Glasswing, has been merged into an expanded verification program.
What the numbers say, and what they do not
The announcement page gives one figure: an expected true-positive rate above 90 percent, which Anthropic says it will work to improve. The other figures come from SiliconANGLE's coverage, and we did not find them on Anthropic's page.
The wolfSSL project reportedly received 74 reports in early trials, all but two of them valid, and five became CVEs.
In a pre-launch check, experts reportedly reviewed 97 critical and high-severity findings across 48 projects and cleared 85 for disclosure. SiliconANGLE says all but one of the other 12 were real bugs that duplicated known issues or other findings.
Nearly 5,000 reports from a manually reviewed backlog have reportedly gone out so far. That is a count of backlog reports and not of OSS Scanner output.
These measure different things. "Valid" in one project's trial, "cleared for disclosure" in a pre-launch review and an "expected true-positive rate" for an unreviewed stream are not the same measure, so do not treat any of them as a rate you can plan on. A rate above 90 percent still leaves a few reports in ten that are wrong, and the first one you trust is the one that costs you.
Why it matters beyond maintainers
Finding a bug and fixing one are different jobs, and the second is the slow one. We wrote about OpenAI's GPT-6 Cyber on November 7, where Chris Hughes of Noma Security, quoted by CSO, pointed out that OpenAI's own dashboard shows a remediation rate below 25 percent, so faster discovery may widen the gap between finding and fixing. We also covered the SAP study from Onapsis and Recorded Future, which reports the median time from a patch to a first exploit falling from 489 days to 1 to 3 days, which we read as a direction and not a clock. If models find more bugs faster, the people who feel it first are the maintainers who have to read the reports, and then everyone who depends on the code.
A triage plan if you are a maintainer
These are our suggestions, not Anthropic's.
Decide your capacity before you enroll. Pick a number of reports per week you can actually handle, and enroll or decline on that basis. Anthropic says the service is meant for projects that can keep up.
Treat every report as untrusted input. A proof of concept written by a model is code someone else wrote. Run it only in an isolated environment with no credentials, never on a developer machine or in your release pipeline.
Rate severity yourself. Anthropic warns that severity ratings may be wrong.
Do not publish a CVE, or change a public advisory, from an unverified report. Reproduce it first.
Send feedback. Anthropic says it will work to improve accuracy and fix quality. Reports marked wrong or duplicate are how that happens.
Keep a record of how many reports were valid, duplicates and wrong, so you have your own rate and not a headline one.
If you depend on open-source code, expect more fixes and more advisories, sometimes with short notice, and keep your patch cadence measured in days for anything exposed to the internet.
Indicators
None. This is a program announcement, so we added nothing to the feed.
Limits
We read Anthropic's announcement page and SiliconANGLE's and SecurityWeek's coverage. We did not read the OSS Scanner page on Anthropic's research site, which is where enrollment is explained, and we did not test the service or see a report. The 90 percent figure is Anthropic's expectation. The wolfSSL, pre-launch and backlog figures are SiliconANGLE's account and were not on the announcement page. The triage plan is ours, and our relationship with Anthropic is a bias you should weigh.
Five percent of this is wrong somewhere. It always is. The first published results from OSS Scanner reports will show how the numbers hold.
Every indicator in this post is in the feed. Free.
1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.
Comments