Anthropic Says Unreleased Claude Model Built Five Cross-Platform Exploit Chains
Anthropic’s Frontier Red Team reported that an unreleased Claude research model discovered vulnerabilities and assembled five working exploit chains against Chrome, Safari, Android, Windows, and Linux in controlled laboratory tests. Reported results included Chrome sandbox escape, Safari cross-origin data access, data extraction from a physically accessible locked Pixel 9, Windows remote denial of service, and unauthenticated remote root access on Linux under specific conditions. Anthropic said all findings were disclosed to the relevant vendors and patched; no real-world attacks involving these demonstrations were reported. Apple and Google acknowledgments corroborate specific vulnerability discoveries, but the model’s claimed autonomy and contribution to each exploit chain remain independently unverified. Reporting also identified unresolved discrepancies between Anthropic’s Windows and Linux exploit descriptions and public CVE records, alongside incomplete corroboration of the claimed Chrome CVE total in Google’s release credits. Security teams should prioritize current browser, operating-system, and Linux kernel updates while distinguishing the demonstrated laboratory outcomes from unverified claims about autonomous exploitation.. Fonte: https://mallory.ai/stories/01a181c8-65b6-7463-b8b4-f6c64494db87
Itagora tiene in archivio titolo e sommario per la ricerca interna. Il testo completo resta sulla fonte. Il link in uscita non invia referrer.