The Record

Whitepapers and long technical documents, read as evidence

ModelsThe Record

NIST AI 800-5 synthesis finds agent threats novel but cybersecurity principles adaptable

Published May 18, 2026; summarizes CAISI RFI NIST-2025-0035 (comments due March 9, 2026).

TLDR

NIST Trustworthy AI 800-5 summarizes RFI responses on AI agent security. Commenters widely agreed agents pose novel threats that hinder adoption and that core cybersecurity practice must adapt—not merely be reused. Government roles cited: implementation guidance, information-sharing, and standards. COSAiS SP 800-53 overlays remain forthcoming; prescriptive control IDs are UNKNOWN until NISTIR 8605D drafts.

Compute & powerThe Record

DeepSeek-V4 paper reports 27% FLOPs and 10% KV cache at 1M tokens versus V3.2

arXiv:2606.19348 (April 26, 2026); MIT checkpoints on HuggingFace; CSA/HCA hybrid attention.

TLDR

DeepSeek-V4-Pro and V4-Flash use interleaved Compressed Sparse Attention and Heavily Compressed Attention for 1M-token contexts. At 1M tokens the paper reports V4-Pro at 27% single-token FLOPs and 10% KV cache versus DeepSeek-V3.2; V4-Flash at ~10% FLOPs and ~7% KV. MIT weights on HuggingFace. Dollar cost per million agent tokens on your stack is UNKNOWN until benchmarked.