Making frontier cybersecurity capabilities available to defenders

• Claude Code Security reasons like a human researcher, not pattern-matching.
  • Traces data flow, understands component interactions, flags business logic flaws.
  • Multi-stage verification reduces false positives; severity ratings assigned.
• Found 500+ previously undetected vulnerabilities in production open-source software.
  • 1+ year of research; partnership with Pacific Northwest National Laboratory.
• Powered by Claude Opus 4.6; dashboard for human review before any patch applied.
• Traditional tools miss complex issues; security teams face backlogs and understaffing.
• Limited preview: Enterprise + Team customers; fast-track for OSS maintainers.

X discourse

• @tenobrus: “Anthropic has capacity to attack superpowers with zero-days; government defenses lag frontier models.” (5466 likes)
• @FurkanGozukara: “US government intentionally dumbed down cyber defenses at DHS, DOJ, Pentagon under Trump priority.” (1958 likes)
• @wallstengine: “OpenAI plans staggered rollout of cybersecurity model to small group, mirroring Anthropic’s restricted Mythos release.” (548 likes)
• @carnot_cyclist: “Better if capability widely available/open sourced long-term; diffusion prevents power centralization.” (242 likes)
• @TheRundownAI: “OpenAI launches GPT-5.4-Cyber for cybersecurity defense, giving vetted defenders frontier capabilities without guardrails.” (17 likes)
• @gagansaluja08: “Defensive angle underrated: frontier model sweeps vulnerabilities, changes parity. Anthropic right to give defenders first.” (0 likes)

· ** · Read on anthropic.com