Anthropic says an unreleased research model found and chained critical bugs into working exploits against Chrome, Safari, Android, Windows and Linux; vendors patched after coordinated disclosure
Anthropic's Frontier Red Team published on Oct 27 five lab attack chains built with an unreleased internal research model. Anthropic says the work needed only initial prompting and little human involvement, reported every flaw under coordinated vulnerability disclosure, and that none of the PoCs were seen in real attacks. How much of each chain the model built alone cannot be independently checked; what can be checked are the vendor patches and credits.
- Chrome (one-click RCE outside the sandbox on stock macOS Chrome): four bugs chained — V8 heap corruption, a V8 sandbox escape, a CacheStorage race that read privileged browser-process memory, and a FedCM use-after-free for sandbox escape. Anthropic says Google assigned and fixed 62 CVEs through Chrome 156 (Oct 20), all critical or high, including 13 sandbox escapes. Google's release notes credit "Xinyang Ge of Anthropic, assisted by Claude" on several Chrome 156 fixes.
- Safari (cross-origin data read): a data-only chain that disables the same-origin policy so a malicious page can read other signed-in origins. Fix: iOS / iPadOS 26.5.2 (Apple advisory).
- Android (locked-phone data exfil over USB): three flaws chained on a Pixel 9 to pull photos and other data without unlocking. Anthropic says Google fixed it; no public CVE called out in the post.
- Windows (remote DoS via HTTP/3 / QUIC): kernel out-of-bounds read, CVE-2026-33096, patched April 2026. Can crash servers that answer HTTP/3 and keep them down by repeating.
- Linux (zero-click root RCE over multipath TCP): race-induced UAF, CVE-2026-98432, patched upstream Oct 15 and backported to maintained stables. Anthropic's PoC took ~48 minutes to leak a kernel pointer and ~5 more to a root shell.
Defender takeaway (Anthropic's): scan your own code and ship PR-time checks; Anthropic also points to its public Claude Code security-guidance plugin and an expanded Cyber Verification Program. Treat this as a capability disclosure from the vendor that built the model, not as evidence of in-the-wild agent attacks on these stacks.
Sources: Anthropic Frontier Red Team: Discovering and exploiting vulnerabilities with Claude (Oct 27) · Chrome Releases: Stable Channel Update for Desktop, Chrome 156 (Oct 20) · Apple: About the security content of iOS 26.5.2 / iPadOS 26.5.2 · Tenable: CVE-2026-98432 · The Hacker News: Anthropic's Unreleased AI Model Chains Flaws Into Working Exploits (Oct 28)