Cryptographic provenance and output watermarking →
In its Oct 19, 2026 Frontier Safety Roadmap Moonshot update, Anthropic renamed 'provable inference' to 'inference verification', rolled out cryptographic signing across all internal token traffic (catching a rare benign integrity glitch), and verified production outputs against committed weights via trusted-server recomputation. On Nov 2, 2026 Anthropic open-sourced FlashProve, a Lean-verified zkSNARK proving 4-bit Llama 4 Maverick (400B MoE) at 110 tok/s on one H100 (18.78 MB proofs, 17s CPU verify), though ~70-95x GPU overhead and a 1,024-token greedy integer limit keep external API text unattested. anthropic.com/research/provable-inference ↗
Anti-sycophancy and pushback post-training →
OpenAI's Oct 29, 2026 Model Spec update elevates Honesty and Transparency (anti-sycophancy, anti-flattery, and multi-perspective balance) to User+1 priority and bars silently relaxing factual accuracy for inferred user preferences, paired with adversarial multi-turn user simulations in the Sept 3, 2026 GPT-6 Astra System Card. openai.com/index/updated-model-spec ↗