Who watches the watchers? — Friday, 9 October 2026
-
Free now, we’ll see later
On the same day Anthropic rewrote its rules for deceptive agents and put Claude beside CrowdStrike inside power plants, it made its strongest models free for open-source scans. A generous Thursday, perhaps. Also a rather clever one
Read the leader →Thursday was a busy day at Anthropic. First came a 2026 Usage Policy refresh, effective 12th November. Bans on fake accounts and fabricated news now sit in one section, Do Not Engage in Deceptive Campaigns or Artificial Activity. The elections rules are renamed Do Not Undermine Democratic Processes. The blanket ban on personalised campaign targeting is gone (it caught nonprofits writing ballot notices in other languages; deceptive targeting remains banned elsewhere). And Claude gets spelled-out stop-buttons when it is wired to hardware that can injure. Most of this, Anthropic says, restates enforcement already under way. Then came the Anthropic Cyber Mission. Its Critical Infrastructure Defense Program (CIDP) puts frontier Claude models, on-site engineers and threat research into the hands of the firms operators already trust with operational technology: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. The argument is fair enough. You cannot simply switch off a power grid, a water plant or a transport network to patch it, and…
-
Fired for talking?
Three fired OpenAI safety researchers say ordinary outside collaboration is now punishable; the company insists on misconduct. Either way, the people paid to watch the models are now watching their backs
Read the leader →Jasmine Wang, Tomek Korbak and Mikita Balesni, the three safety researchers OpenAI fired last week, published an open letter on Thursday. They deny the company’s claim that they mishandled sensitive information, and warn that colleagues are now “afraid to speak.” AI safety work, they say, depends on “close collaboration with…
-
The colleague who needs no salary
Google’s Gemini agent gets its own email and audit trail, for businesses first. The personal-agent race is being won at the office, where nobody asks the new hire too many questions
Read the leader →At a Google Cloud event on Thursday, Gemini crossed from answering questions to “getting things done.” The new enterprise agent takes objectives, not just instructions: it plans work, loads skills, and connects to Workspace, Microsoft 365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Postgres, Snowflake and any Model Context Protocol server…
Labs
Anthropic’s Cyber Mission puts frontier Claude, on-site engineers and threat research beside CrowdStrike, Palo Alto Networks, Rockwell and others to defend grids, water and transport. A free, opt-in OSS Scanner sends unreviewed model-generated vulnerability reports to open-source projects; unreviewed, that is, no human reads them first. anthropic.com
Products
Google’s enterprise agent gets its own Workspace identity, email address and audit trail, and connects to Slack, Jira, M365, data warehouses and MCP servers. Businesses get it first; the rest of us wait. techcrunch.com
Business & funding
The Financial Times says OpenAI told investors annualised revenue is “approaching $50 billion”, about $20bn below a circulating $70bn figure that tried to match Anthropic’s partner-inclusive counting. That is rather a large rounding error. techcrunch.com
Policy & society
A 2026 Usage Policy, effective 12th November, folds bans on fake accounts and fabricated news into one section on deceptive campaigns and drops the blanket ban on personalised campaign targeting. Models wired to hardware that can injure must now have a stop-button and a safe state; good to see that written down somewhere. anthropic.com
- Long-WAM. Memory helps a robot only if it can imagine the future
NVIDIA’s Long-WAM shows longer visual history lifts success when video pretraining is causal—and runs in 107 ms on a consumer GPU
- Agent Oversight EBG. Not whether the agent finished—whether you can find what it changed
Evidence-Grounded Behavior Graphs improve localization of consequential decisions across eight models, and help Codex disclose silent changes
- Recursive Game Creator. Playable is not the same as fun
A Designer–Builder–Player–Reviewer harness scores player experience, lifting GameCraft to 77.89 and GameASG strict success to 53.2%
- Beijing will not pace with you
SemiAnalysis
A census of 857 Chinese model releases: only 3.6% ever published a safety result.
- Probability without a novel
Timothy B. Lee · Understanding AI
The clearest plain-English guide to TypeSafe’s probability-only Jev model and why rivals are copying it.
- What the agents missed in the UV sky
Brice Ménard · Anthropic
A candid field note on Claude Science, including the artefacts the agents missed and a human caught.
- False fronts, Category 5
OpenAI
The richest primary-source case study yet of AI-assisted influence operations, invented journalists and all.
- When the drop is process, not a model
Zvi Mowshowitz · Don't Worry About the Vase
A dense weekly map of a week defined by math standards and safety firings; follow the receipts, and bring coffee.
Who’s gaining ground
Big gets bigger; capability still leaks out. For now.
Concentrating
Score by issue
Score from −5 (spreading) to +5 (concentrating)