Anthropic has launched OSS Scanner, a free, opt-in vulnerability scanner that runs its strongest models, including Claude Mythos, against critical open-source projects. The pitch is speed: reports go straight to maintainers without human review, a trade-off that delivers findings faster but shifts the triage work onto project teams.
Anthropic OSS Scanner (Credits: Anthropic)
Anthropic OSS Scanner: Free AI Vulnerability Scanner for Open Source
The scanner is run jointly by Anthropic’s Frontier Red Team and Safeguards Research team, and its design borrows from Google’s OSS-Fuzz while swapping fuzzers for language-model agents. Maintainers enrol by opening a pull request on Anthropic’s oss-scanner repository, supplying a repository, a primary contact and a Dockerfile; audits then run in an isolated virtual machine with no internet access.
Each report contains a self-contained reproducer and an explanation of the bug, plus a bisection to the commit that introduced it where possible, and a candidate patch when one is available. Agents in the pipeline double-check bugs, run root-cause analysis and propose fixes before anything is sent out.
Scale so far is striking. Anthropic says 114 projects were accepted from about 650 applications as of October 19, and just 11 days after the October 8 launch it reported 77 critical and 803 high model-rated findings. The programme grew out of a review bottleneck: over six months, Anthropic’s models surfaced more than 29,000 candidate vulnerabilities, human reviewers triaged roughly 6,000 of them, and nearly 5,000 more reports went directly to maintainers who asked for everything, validated or not.
OSS Scanner Results: 88% of Audited Findings Met the Disclosure Bar
The early accuracy numbers look strong. Anthropic asked its penetration testers to check 97 critical and high findings across 48 projects, and 85 of them, or 88%, met the bar for its coordinated vulnerability disclosure process. Only one of the 97 was a false positive, though 11 duplicated known issues or other scan findings.
Maintainer experience points the same way. WolfSSL says 72 of the 74 reports it received were valid, and five became CVEs. Curl patched CVE-2026-92392, a high-severity HTTP/2 upload use-after-free affecting curl 8.12.0 to 8.22.0, fixed in 8.23.0, released October 14, 2026.
The caveats matter too. There is no human review before delivery, and maintainers report inflated severity ratings alongside the duplicates. Anthropic links the timing to benchmark progress: on CyberGym, language models went from finding under 20% of vulnerabilities to over 85%. Its broader disclosure dashboard lists 6,157 disclosed vulnerabilities across 591 projects as of October 2, with 516 patched.
For open-source maintainers, the offer is hard to ignore: a free audit by frontier models, with Anthropic covering the full cost. The open question is whether project teams can absorb the flood of unreviewed reports, duplicates included, without the scanner becoming another inbox to triage.
Also Read: Meet Anthropic’s Sandbox Runtime: OS-Level Sandboxing for AI Agents, No Containers Needed | Anthropic IPO November Deadline Looms: Will the Claude Maker File Its S-1 in Time?
Source: MarkTechPost — read the full story there.
