The frontier is writing the handshakes — Aug 28

Four Aug. 27 signals put protocols, coalitions, sealed evals, and a court order between frontier AI and the outside world.

Anthropic opened Model Hardware Standard. The research preview lets agents operate microscopes, liquid handlers, robotic arms, and other programmable lab and factory devices through a shared, model-agnostic driver. It began with HHMI Janelia; early partners include Genentech, QuEra, AWS Strands Robots, and Hugging Face LeRobot. Anthropic plans to open-source it after safety work.1

OpenAI led a 100+ company cyber-defense letter. Signatories including Anthropic, Microsoft, Alphabet, Amazon, Cloudflare, CrowdStrike, IBM, Mastercard, Oracle, and Visa warn that AI-enabled attacks on hospitals, water plants, and internet infrastructure will get far more widespread in the coming months, and call for a society-wide defensive surge.23

DeepMind piloted double-blind frontier evals. Working with Singapore’s AI Safety Institute, OpenMined, AVERI, and MLCommons, Google DeepMind tested a Gemini Flash Lite model inside Google Cloud Confidential Computing so evaluators never see weights and Google never sees reserved AILuminate prompts.45

A federal judge blocked the Pentagon’s Anthropic blacklist. U.S. District Judge Rita Lin voided the Defense Department’s national-security supply-chain risk designation, which followed Anthropic’s refusal to allow Claude for U.S. surveillance and autonomous weapons. Lin called the label illegal and baseless.6