In plain terms
Two-tier AI disclosure architecture. Public tier (Opus 4.7, GPT-class commercial models, LLaMA open weights) is heavily AI safety training (RLHF)-constrained and benchmark-degraded.
Two-tier AI disclosure architecture. Public tier (Opus 4.7, GPT-class commercial models, LLaMA open weights) is heavily AI safety training (RLHF)-constrained and benchmark-degraded. Gated tier (Claude Mythos Preview via Glasswing consortium; adversary-side DeepSeek V4 to Huawei/Cambricon) delivers autonomous frontier capability — zero-day exploitation, 73%+ expert CTF success, full TLO 32-step attack simulation completion. The public/gated split mirrors the same architecture already mapped at the compute layer (the nuclear-AI project (Genesis Mission)), information layer (the bot-dominated internet), and financial layer (BlackRock's Aladdin algorithm). Predicted for 2027, fired April 2026 — ~10 months early.