Claude chained real zero-days: Linux root, Chrome RCE, Pixel bypass
An unreleased Anthropic model built end-to-end exploits across five OSes with near-zero prompting. Verified cyber access expands; Claude Code 2.1.312 ships too.
Copy markdown
Five working exploit chains, near-zero prompting
Anthropic's Frontier Red Team had an unreleased internal model build end-to-end exploits across five platforms: a one-click Chrome RCE that ran Calculator on stock macOS, a Safari same-origin bypass, a Pixel 9 lock-screen photo grab, a Windows QUIC kernel DoS (CVE-2026-33096), and a zero-click Linux MPTCP root RCE (CVE-2026-98432) that reached command execution in under an hour. All were filed under coordinated disclosure.
Verified security pros get less-restricted models
Anthropic expanded its Cyber Verification Program, which hands qualifying security professionals reduced blocking classifiers on Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 for offensive work. Not verified? The defensive move today is wiring Anthropic's open-source Claude Code security reviewer into your pipeline as a GitHub Action that scans pull requests.
Patch list: Chrome 156 and two kernel CVEs
Everything demoed is already fixed upstream, so update. Chrome shipped 62 patches through version 156 (13 sandbox escapes, 21 renderer-compromise bugs), and the Linux and Windows kernel CVEs are patched too — Anthropic estimates the Windows QUIC bug exposed several million internet-facing servers.
Claude Code 2.1.312 lands
Cloud routines can now pick an effort level in the model picker, and fullscreen mode gets click-to-send for queued messages. Note the renames and caps: the rolling window is now the '5-hour limit' (was 'session limit'), and hooks are capped at 24 running per session, sub-agents included.