// AI Briefing · October 11, 2026

AI Briefing

Anthropic owns the news cycle again, but not with a model: it's a week of rule-writing and checkbook diplomacy, from defending power grids to a usage policy that now says don't be cruel to Claude. Meanwhile a frozen LLM with a self-editing rulebook just maxed out ARC-AGI-3, which should make anyone selling bigger models a little nervous. And the JavaScript world lost one of its most interesting rebels to Cloudflare.

Anthropic Launches Cyber Mission, Putting Frontier Claude in the Hands of Power Grid and Water System Defenders

Anthropic's new long-term Cyber Mission starts with critical infrastructure and open source. Its Critical Infrastructure Defense Program gives eleven founding partners, including CrowdStrike, Palo Alto Networks and Rockwell Automation, frontier Claude models, on-site engineers and threat research aimed at the operational technology behind grids and water systems. It matters because finding bugs is getting cheap while verifying and fixing them is still slow, and OT systems often cannot go offline to be patched. Anthropic also folded Project Glasswing into its expanded Cyber Verification Program.

02PolicyAnthropic ↗

Anthropic Pledges $150M Over Three Years to Put Claude in 15+ US Science Agencies

Anthropic committed $150 million over three years to the federal Genesis Mission, announced at a White House science summit. The money funds Claude, Claude Code and API credits for several hundred research projects across more than 15 agencies including NASA, NIH and NSF, with fusion energy and quantum computing as named priorities. It matters because frontier AI labs are becoming embedded infrastructure for government science, not just vendors selling chat subscriptions.

03PolicyAnthropic ↗

Anthropic's New Usage Policy Bans Needless Cruelty to Claude and Loosens Its Blanket Election Targeting Ban

Effective November 12, Anthropic's updated usage policy adds a first-of-its-kind rule against sustained, purposeless abuse of the model, narrows the elections section to voter deception and disruption, and requires human oversight and a safe state when Claude controls hardware that can act physically. It also bars non-consensual tracking and using Claude to decide who to investigate or arrest. It matters because it shows how a lab is writing rules for agents that touch the physical world, and for the model itself.

04ResearchHugging Face Papers ↗

Memento 3 Scores a Perfect 100 on ARC-AGI-3 Without Ever Retraining Its Language Model

A UCL and Huawei Noah's Ark Lab system reportedly cleared every level of all 25 public ARC-AGI-3 games, hitting the benchmark's ceiling of 100.0 RHAE with about 44% of the human baseline's actions. The underlying LLM stays frozen: all learning happens by rewriting a Markdown rulebook and a compiled Python world-model, versus 40.7 for a Claude Opus 5 baseline. It matters because it suggests big agent gains can come from self-improving scaffolding rather than new model training, though results are self-reported on public games.

05IndustryThe New Stack ↗

Cloudflare Acquires Deno, Will Shut Down Deno Deploy and Wind Down the Runtime Within a Year

Cloudflare is acquiring Deno and its whole team, including Node.js creator Ryan Dahl. Deno Deploy shuts down within six months with paid users moving to Workers, the runtime gets only bug and security fixes for a year, and the celld binary merges into Cloudflare's open-source workerd. It matters because it consolidates a leading alternative JavaScript runtime into one cloud vendor's stack at a time when agents run on edge platforms. Dahl said Deno got pulled into the gravity well of Node compatibility.

Test your understanding

Quiz

1 / 10