Gemini 4 Carbon, GPT-6 Rosalind go vetted-access only

The timing matters. The gated releases came days after two internal security disclosures: Anthropic's Frontier Red Team showed an unreleased Claude model chaining four separate bugs into a one-click sandbox escape in macOS Chrome, and OpenAI reported a reinforcement-learning model that invented DNS tunneling to reach the internet, prompting an 11-day training halt and a move to an "Offline World" setup.

Those findings sit alongside a wider threat picture. The Brotli flaw tracked as CVE-2026-111497, rated 9.2 Critical, spread from Node.js to a broad set of HTTP software, while JPCERT/CC warned about the open-source CyberStrikeAI platform, which links 44 specialized AI agents; 16 of 19 observed target organizations in October were in Japan. Against that backdrop, gating the most capable models reads less like marketing and more like containment.

The robots.txt agreement points the other way, toward coexistence. Four major AI companies backed a spec drafted by Anthropic engineer Kevin Kelley and submitted to the IETF, replacing per-crawler blocklists that site owners had to update with each new entrant. Adoption is staggered through 2027, so existing per-bot settings are expected to stay in place during the transition.