LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
← Voltar para a leitura
NotíciasMaterial2 min de leituraEditorial DisconfirmAI

Anthropic Launches OSS Scanner: Free Periodic AI Vulnerability Scans for Open Source, Reports Unverified

Anthropic launched OSS Scanner on Oct 8, offering eligible open-source projects free periodic vulnerability scans by its strongest models. The service provides automated, unverified reports that may contain false positives, accelerating discovery but shifting triage burden to maintainers.

2026-10-08 08:00 GMT+8

US-based AI company Anthropic officially launched OSS Scanner on October 8, 2026, a voluntary vulnerability scanning service for the open-source ecosystem. Eligible projects can receive free, periodic security scans executed by Anthropic’s strongest models, including Claude Mythos. This initiative stems from Anthropic’s experience using Claude to discover vulnerabilities during "Project Glasswing," aiming to address the delays caused by human review bottlenecks in traditional Coordinated Vulnerability Disclosure (CVD) processes.

Unlike previous CVD reports which underwent human verification, OSS Scanner outputs are entirely model-generated without human review or triage. Anthropic explicitly states that while this "fast track" enables more frequent and rapid scanning, it also means reports may be incorrect or invalid. Each report includes a proof-of-concept (PoC) and, where feasible, a candidate patch. For maintainers, this implies they must independently assess the validity and priority of each report rather than relying on vendor pre-screening.

Eligibility criteria mirror those of Google’s OSS-Fuzz, targeting mature projects with critical impact on infrastructure and user security. Factors include remote attack risk, number of users/downstream dependencies, and OpenSSF criticality scores. Brand-new projects and most forks are typically ineligible. Core maintainers must register by submitting a Pull Request to a GitHub repository, providing a Dockerfile capable of offline builds to support sandboxed auditing.

Anthropic noted that over the past six months, its models discovered more than 29,000 candidate vulnerabilities, but only about 6,000 could be manually reviewed. OSS Scanner aims to release this previously unprocessed discovery capability. Despite the risk of false positives, early feedback suggests high signal quality for some recipients; for instance, wolfSSL reported that all but two of 74 received reports were valid, with five becoming CVEs. However, Anthropic acknowledges that severity ratings may be inflated and that unverified findings do not adhere to the standard 90-day coordinated disclosure period.

Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

Fontes

  • anthropic.com↗
  • red.anthropic.comFrontier Red Team Research↗
Continuar pesquisando →