What should a developer do when its models find vulnerabilities faster than people can check them? This post, published on 8 October 2026 by Anthropic's Frontier Red Team with no individual byline, announces OSS Scanner: an opt-in service that sends enrolled open-source projects vulnerability reports that are "fully model-generated, without human review or triage" ("Our vulnerability scanner"). The page's heading reads "Launching an opt-in vulnerability-finding service for open-source software." Three documents together make up the version of record: the post (about 1,000 words), the service's FAQ at red.anthropic.com/oss-scanner (changelog: launched 8 October, revised 13 October 2026), and Anthropic's coordinated vulnerability disclosure (CVD) dashboard, whose dated snapshots carry the program's counts. The FAQ says the service is a collaboration between the Frontier Red Team and the Safeguards Research team.
Why the post says a fast track is needed
The post's case rests on a backlog (introduction):
- Capability. On CyberGym, an academic vulnerability-finding benchmark, models went "from finding under 20% of vulnerabilities at the beginning of last year to finding over 85% this year." The post gives no model names or source for these figures beyond a link to the benchmark paper.
- Backlog. Over six months, Anthropic's models found "over 29,000 candidate vulnerabilities," of which only about 6,000 had been manually reviewed and triaged: "we remain bottlenecked on our human capacity to validate these findings."
- Maintainer requests. Maintainers who received the first reports increasingly asked for everything. Anthropic had sent "nearly 5,000 reports directly to maintainers after they asked to receive everything we had—even if it wasn't validated."
- Speed. Because exploits "can now be developed in minutes" (linking Anthropic's post on N-day exploits), faster fixes matter.
Human-verified reports continue through the existing CVD process, "especially for projects without the resourcing to triage reports themselves." OSS Scanner is an additional, optional route.
The dashboard behind these numbers. The CVD dashboard's snapshot of 2 October 2026, six days before the post, gives 29,439 candidates, 6,123 findings reviewed by external security firms and 5,674 confirmed valid (92.7%), with 4,824 findings "Reported direct to maintainers by Anthropic" marked "May contain false positives." Matching these to the post's rounded figures is this note's inference. The dashboard's direct-disclosure category also covers findings sent directly for other reasons, not only on request. The snapshot of 2 November 2026 lowers the direct count to 3,225 and total reports from 6,157 to 5,087. The dashboard's About page, in its "Changes between updates" section, explains that "many reported findings were duplicates of our other findings," which were marked as such and no longer counted, along with some withdrawn findings. The post's "nearly 5,000" predates that correction.
What the service sends and to whom
From the post and the FAQ:
- Models. Reports come from "our strongest models (including Claude Mythos)" (post). The FAQ adds that Anthropic uses several harnesses, including "token-hungry and experimental" ones, and covers the full cost.
- Report contents. Each report has a self-contained reproducer, an explanation of the vulnerability with a bisection to when it was introduced "where possible," and a candidate patch "when available" (post). The pipeline includes agents that double-check bugs, propose patches and analyze root causes (FAQ).
- Enrollment. Core maintainers open a pull request to a public GitHub repository with a configuration file and a Dockerfile that builds the project for scanning with no internet access (FAQ). Eligibility uses criteria similar to OSS-Fuzz's: projects with "critical impact on infrastructure and user security," decided case by case.
- Disclosure. Unvalidated findings carry no 90-day disclosure deadline, because they "may contain false positives." A report that Anthropic later validates through the CVD program may be disclosed 90 days after the maintainer is told it was validated (FAQ).
- Scope controls. In the FAQ as revised on 13 October (the changelog says it "Added further explanation" of how reports reach only a project's maintainers), a person reviews each sign-up to confirm it comes from a core maintainer, the security contact is verified, and the harness checks that each report concerns the project rather than its dependencies.
The pilot results
The post reports a pilot with "dozens of open source projects" over "the last several weeks," which produced "hundreds of bug reports," including several vulnerabilities that Anthropic chained into unauthenticated remote code execution. It gives one quantitative check ("Our vulnerability scanner"):
- Expert review. The penetration testers who review CVD findings checked 97 critical and high-severity findings from an early scanner version, across 48 projects. 85 (88%) met the bar for the CVD process. Of the other 12, 11 were real but duplicated known issues or other scan findings, and one was invalid. The dashboard defines its own true-positive rate to include duplicates; on that definition, 96 of 97 would count (this note's arithmetic).
- Maintainer feedback. Maintainers "have seldom told us a high or critical finding was invalid," though some said severity ratings "can be inflated" or that the scanner "misunderstood the project's threat model." No counts are given.
Five maintainers are quoted, from PostgreSQL, OpenSSL Corporation, wolfSSL, HotCRP and curl. The wolfSSL quotation carries the only other count: "of the 74 reports we received, all but two were valid, and five became CVEs." These are testimonials chosen by Anthropic.
The dashboard's severity comparison bears on the "inflated" remark. In the 2 November snapshot, Claude's initial severity matched the external firm's rating exactly for 88.0% of 1,868 triaged findings, and was within one band for 97.9%. In the 2 October snapshot the figures were 83.0% and 97.1% of 1,337. The dashboard says firms' ratings "tend to be lower" because they include project-specific context. These figures concern the human-triaged CVD findings, not OSS Scanner's unreviewed reports.
The written threat model
The service lets maintainers steer the scanner with a document. Projects may supply a threat-model file, by default .oss-scanner/threat_model.md, with "no required format or sections" (FAQ). The FAQ suggests it describe what code to test, which inputs are adversarial and what is out of scope, a severity rubric, report and patch format, and deduplication granularity. Without it, the scanner "will make guesses at what it thinks is the best way to handle any corner cases." Maintainers may edit the file between scans "to alter the report format." Neither document reports whether supplying or changing this file alters the scanner's findings, severity ratings or false-positive rate.
Limitations
- Self-reported precision. The 97-finding check was done by the penetration testers who review Anthropic's own CVD findings, on an early version and on critical and high findings only. No precision is reported for medium and low findings or for the launched version.
- Unreviewed by design. The post says "it is possible reports will be incorrect or invalid," and "We can't guarantee the scanner will be perfect."
- Unlocated capability claim. The CyberGym figures name no model or source.
- Moving counts. The backlog numbers come from a dashboard that has since revised related totals downward for duplicates.
- Selected testimonials. Maintainer quotations are chosen by Anthropic.
Research questions
- What is OSS Scanner's precision at each severity? Maintainer-reported outcomes for every report, by severity, would replace the single 97-finding check.
- Does a maintainer's threat-model file change what the scanner reports? Scanning the same project with and without the file, or with different rubrics, and comparing findings and severity ratings would show whether the written guidance is followed.
- Do unreviewed reports burden maintainers? Time spent per report, and the share of projects that pause or leave the service, would test the trade-off the post accepts.
- How fast are fast-tracked findings fixed? Comparing time to patch for OSS Scanner reports with CVD-triaged reports in the same projects would test the speed argument.
- How many duplicates do model-generated reports contain? The dashboard's November correction suggests the rate is material; publishing it would help maintainers judge volume.
- How do benchmark scores relate to open-source yield? Reporting the CyberGym score of each scanning model beside its confirmed findings would connect the post's two kinds of evidence.
How it connects
- The post links Assessing Claude Mythos Preview’s cybersecurity capabilities, Evaluating and mitigating the growing risk of LLM-discovered 0-days and Measuring LLMs’ impact on N-day exploits for the capability background.
- Using Claude to reduce the attack surface of code (October 2026) describes a skill being integrated into OSS Scanner, and Discovering and exploiting vulnerabilities in critical, real-world software with Claude (October 2026) opens by linking this post.
- Partnering with Mozilla to improve Firefox's security (March 2026) describes how Anthropic made earlier reports trustworthy for maintainers; this note connects it as the human-verified approach that OSS Scanner supplements.
- Building AI for cyber defenders (September 2025) reports CyberGym scores from 2025, including 28.9% for Claude Sonnet 4.5 under the leaderboard's cost limit; the connection is this note's.
- The Mythos model family is documented in the Claude Mythos Preview system card and later Mythos cards.
References
- Anthropic Frontier Red Team. (2026, October 8). Launching an opt-in vulnerability-finding service for open-source software. https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source record
- Anthropic Frontier Red Team. OSS Scanner FAQ (changelog entries 2026-10-08 and 2026-10-13). https://red.anthropic.com/oss-scanner record
- Anthropic Frontier Red Team. Anthropic's coordinated vulnerability disclosure dashboard. Snapshots of 2026-10-02 19:47 UTC and 2026-11-02 19:30 UTC, and the About page. https://red.anthropic.com/2026/cvd/ record
- OSS Scanner enrollment repository. https://github.com/anthropics/oss-scanner record
- Wang, Z., et al. (2025). CyberGym: Evaluating AI Agents' Cybersecurity Capabilities with Real-World Vulnerabilities at Scale. arXiv:2506.02548. Cited in the post; not read for this note. record · 2 pages