Reading
Anthropic's Free Bug Hunter Skips Humans. Who Cleans Up?
Anthropic's OSS Scanner is free, but no human reads its bug reports before they reach maintainers. Here's the catch, and who should sign up.
Prices last read 2026-10-22
Short answer
The catch is that no human checks the reports first. Under the terms, checking them is your job.On October 8, 2026, Anthropic launched the Anthropic Cyber Mission. One part is OSS Scanner, a free, opt-in service that runs periodic scans of enrolled open-source projects using Anthropic's most capable models. Each report includes a proof of concept, an explanation and, where available, a suggested fix. Reports are sent without human review. Anthropic says it expects a true-positive rate above 90%, and its terms put the job of reviewing every report and patch on the maintainer.Source: Anthropic: Introducing the Anthropic Cyber Mission · 2026-10-22; OSS Scanner Terms & Conditions · 2026-10-22
What Anthropic actually launched
The Cyber Mission starts with two programs. The Critical Infrastructure Defense Program brings frontier Claude models, on-site engineers and threat research to companies that secure operational technology. Its 11 founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. The second program is OSS Scanner, which Anthropic says was inspired by Google's OSS-Fuzz. Anthropic also says it funded organizations behind widely used open-source code, naming the Python Software Foundation, Alpha-Omega and OpenSSF (through the Linux Foundation), and the Apache Software Foundation. It does not give dollar amounts.
Source: Anthropic: Introducing the Anthropic Cyber Mission · 2026-10-22
Which model does the scanning? Anthropic's post only says its "most capable models." TBreak reports that these include Claude Mythos. On accuracy, SiliconANGLE reports that experts checked 97 findings from an early version across 48 projects and cleared 85 for disclosure, and that all but one of the other 12 were real bugs. The same report says wolfSSL found all but two of its 74 reports valid.
Source: Anthropic: Introducing the Anthropic Cyber Mission · 2026-10-22; TBreak: Anthropic OSS Scanner free security scans · 2026-10-22; SiliconANGLE: Anthropic launches critical infrastructure program and free OSS Scanner · 2026-10-22
The catch, in the scanner's own fine print
Reports may be incomplete and inaccurate.
OSS Scanner Terms & Conditions, section 4a (read 2026-10-22)Source: OSS Scanner Terms & Conditions · 2026-10-22
That line matters because the service is provided "as is," and section 4c says the participant is responsible for reviewing each report and any patch it proposes. The GitHub README is just as blunt: reports are model-generated and not reviewed by a human. Anthropic says the trade-off is speed, and it warns that some reports could get details like severity wrong. Because these findings have not been validated, they carry no 90-day disclosure deadline and Anthropic says it will not make them public. If a finding is later validated through Anthropic's coordinated disclosure program, it may be disclosed 90 days after maintainers are notified.
Source: OSS Scanner Terms & Conditions · 2026-10-22; GitHub: anthropics/oss-scanner README · 2026-10-22; Anthropic Red: OSS Scanner · 2026-10-22
Who is affected
Not every project can join. Anthropic accepts "established projects that have a critical impact on infrastructure and user security" and decides case by case. It weighs factors such as age, update frequency, OpenSSF criticality score and adoption. New projects and most forks generally don't qualify. Under the terms, you must be the lead maintainer or be authorized by one. Anthropic says the service is meant for projects that can keep up with the findings. Projects that can't will keep getting human-verified disclosures under Anthropic's existing coordinated disclosure policy.
Source: Anthropic Red: OSS Scanner · 2026-10-22; OSS Scanner Terms & Conditions · 2026-10-22; Anthropic: Introducing the Anthropic Cyber Mission · 2026-10-22
- Open an enrollment PRAdd a projects/<name>/project.yaml file to the anthropics/oss-scanner repository. It needs the repo URL, a primary_contact address and a Dockerfile that builds the project. Anthropic reviews and merges enrollment PRs and checks that you are a core maintainer.
- Use a public-safe addressThe README warns that email addresses in project.yaml are public. Use a security alias, not a personal inbox.
- Make the build work offlineSetup has network access, but the audit runs without internet. Your Dockerfile must fetch everything the build and tests need during setup.
- Write a threat modelThe optional threat_model.md is "strongly recommended." Use it to state your security goals, untrusted entry points and out-of-scope components.
- Know the off switchSet disabled: true to pause reports, or remove your project directory to withdraw. The terms let you pause or end enrollment at any time.
Source: GitHub: anthropics/oss-scanner README · 2026-10-22; Anthropic Red: OSS Scanner · 2026-10-22; OSS Scanner Terms & Conditions · 2026-10-22
Our take: a real gift, priced in maintainer attention
Let's be fair first. Free scans from a frontier model, with a working proof of concept and a draft patch, beat anything most volunteer projects could ever pay for. If the early numbers hold, this is far better than the AI-generated bug spam maintainers have been complaining about for two years. But "no human review" is not a neutral design choice. It moves triage costs from Anthropic's payroll to unpaid maintainers' evenings, and the terms make that explicit. A 90% hit rate sounds great until you remember that the other 10% still lands in someone's inbox. And that rate is a projection, not an independent audit.
The overrated part is the headline: "free." The underrated part is the opt-in design and the off switch. Anthropic didn't force this on anyone, and projects that can't absorb the volume keep the slower, human-checked path. That's the right call. Our advice: well-staffed projects with a real security team should enroll now. Solo maintainers should wait and see what the first batch of reports looks like for projects their size. Either way, treat every report as a lead to verify, not a verdict. And never merge an AI-written patch you haven't read.
Want the model, not the scanner?
Anthropic says maintainers can apply to Claude for Open Source for free Claude Max subscriptions. Security teams can apply to the expanded Cyber Verification Program for defensive access to Claude's cyber capabilities.
Source: Anthropic: Introducing the Anthropic Cyber Mission · 2026-10-22
Before you enroll
- Someone on the project can triage security email within days, not months
- Your primary_contact is a shared security alias you're happy to see published
- Your Dockerfile builds and runs tests with no network at audit time
- You have a threat_model.md that marks out-of-scope components
- Everyone agrees that no AI-proposed patch is merged without human review
Source: GitHub: anthropics/oss-scanner README · 2026-10-22; OSS Scanner Terms & Conditions · 2026-10-22
Optional: compare Claude subscription offers from sellers. This is a seller listing, not official Anthropic access and not an endorsement. OSS Scanner itself is free and needs no subscription.. That link carries our partner code Sk52g, which is how a purchase is attributed to this site. Promotion code UPD3S.
Sources (all read 2026-10-22)