FULL STORY

Anthropic Launches Cyber Mission to Defend Critical Infrastructure

Anthropic announced its long-term Cyber Mission initiative to defend critical infrastructure and open-source software with frontier models, launching the free OSS Scanner, which has already identified some 29,000 candidate vulnerabilities.

2026-10-09 ~ 2026-10-09 · 2 episodes · 10 posts

Episode 1 · Anthropic Launches Cyber Mission to Defend Critical Infrastructure and Open Source (2026-10-09, 7 posts)

On October 9, Anthropic officially announced a long-term safety initiative, the "Anthropic Cyber Mission," aimed at using frontier models to protect the critical systems society depends on, launching with two major programs: the Critical Infrastructure Defense Program (CIDP) and the open-source vulnerability scanning tool OSS Scanner.

Confirmed

  • CIDP targets operational technology and government systems such as power grids, water, and transportation, bringing frontier Claude models, embedded engineers, and the latest threat research to system operators and their security, manufacturing, and technology vendors—helping these long-underresourced defenders counter nation-state threats
  • Inspired by OSS-Fuzz, OSS Scanner provides free periodic vulnerability scans with the strongest models to volunteer open-source projects, helping maintainers get faster bug reports, PoCs, and fixes; the program surfaced 29,000 candidate vulnerabilities in six months
  • On the CyberGym benchmark, LLM vulnerability discovery rates rose from under 20% in early last year to over 85% this year, and the quality of reports maintainers receive has improved markedly from "AI junk reports"
  • Anthropic says Cyber Mission is part of a larger effort to make the systems society relies on safer and more resilient, and it plans to expand and adapt based on lessons learned with public and private sector partners

Why it matters

  • Security research has long been understaffed, especially among critical infrastructure defenders and open-source maintainers; frontier models directly aiding the defensive side is a rare real-world deployment of AI safety capabilities
  • The 85%+ discovery rate on CyberGym shows large models have gone from gimmick to genuinely useful in real vulnerability hunting, validated by the scale of 29,000 candidate vulnerabilities found

Episode 2 · Anthropic Launches OSS Scanner, Uncovering 29,000 Vulnerabilities in Open Source Projects (2026-10-09, 3 posts)

Anthropic has launched OSS Scanner, a free service that uses its latest Claude models to scan volunteer open source projects for vulnerabilities, generating reports with PoCs, explanations, and fix suggestions. It has already identified 29,000 candidate vulnerabilities, of which humans could only review about 6,000.