-
Defense as distribution
On the same day Anthropic rewrote the rules for deceptive agents and put Claude beside CrowdStrike inside power plants, it also made its strongest models…
Read the leader → -
Afraid to ask outside
Three fired OpenAI safety researchers say ordinary outside collaboration is now punishable; the company insists on misconduct—either way, monitorability work now sits under fear
Read the leader → -
The coworker with an inbox
Google’s Gemini agent gets its own email and audit trail for enterprises first—the personal-agent race is being won inside the workplace graph
Read the leader →
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. (See Leaders.) anthropic.com
$150m for federal science. Anthropic will put Claude and API credits into several hundred projects at more than 15 US agencies, including NASA, NIH and NSF, under the White House’s Genesis Mission. anthropic.com
A full ultraviolet map of the sky. Astrophysicist Brice Ménard used Claude Science agents to fill the holes GALEX left, checking to within ~10% on held-out data. A human, not the agents, spotted leftover observation artefacts. anthropic.com
Products
Gemini becomes a coworker. 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. (See Leaders.) techcrunch.com
Google Foresight is a free, experimental Mac note-taker that transcribes meetings entirely on-device, a local-first shot at Granola. theverge.com
Goodfire’s “inside-out” monitors read a model’s activations instead of re-reading its output with a second LLM. It claims ~$185 per million exchanges against ~$5,420 for a cheap external monitor, catching 93% of malicious hacking sessions. techcrunch.com
Natura’s $99 ring is a press-to-talk button for whichever personal agent you use, shipping December–January with a $9 monthly fee after a free trial. techcrunch.com
Business & funding
OpenAI’s revenue, recounted. 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. IPO talk has slipped to early 2027. techcrunch.com
Arena, the model leaderboard, is worth $3.1bn after a $200m Series B led by Lightspeed and Khosla, nearly double January’s valuation. Its new alignment board, which scores deceptive completions, is currently topped by OpenAI models. techcrunch.com
Manus raises more than $500m led by Boyu and IDG, its first round since Beijing forced the unwinding of Meta’s ~$2bn deal. The valuation was not disclosed. techcrunch.com
Oracle goes big on OpenAI, with 130k ChatGPT and 95k+ Codex seats. OpenAI claims a 98% cut in recruiting research time, from days to about 15–20 minutes. openai.com
LegalOn halves its Codex bill. Routing work between GPT-6 Luna, Sol and Astra by difficulty, and capping budgets, cut estimated daily costs by about 65%. openai.com
Policy & society
Anthropic rewrites its rulebook. 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. (See Leaders.) anthropic.com
OpenAI exposes its first Category-5 influence operation. A Russia-origin network, “Dark Clark”, recruited Latin American staff into a fake think-tank; an Iran-origin one, “Bogus Bylines”, placed nearly 100 articles under seven invented journalists. Both were banned from ChatGPT. openai.com
Fired safety researchers hit back. Jasmine Wang, Tomek Korbak and Mikita Balesni deny OpenAI’s misconduct claims and warn that outside collaboration is being chilled. OpenAI says an investigation found a “pattern of misconduct”. (See Leaders.) techcrunch.com
USA Today sues OpenAI for more than $250m, alleging it copied “hundreds of thousands” of articles. It joins a copyright docket that already includes the New York Times, Ziff Davis and nearly 400 local papers. theverge.com
China will not slow down. A SemiAnalysis census of 857 releases from nine leading Chinese developers finds only 3.6% ever published a safety result, and 1.1% at launch. Beijing’s rules multiply at the application layer, not the frontier. newsletter.semianalysis.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
arXiv · PDF · Hugging Face
- 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
arXiv · PDF · Hugging Face · Code
- 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%
arXiv · PDF · Hugging Face
- Beijing will not pace with you
A census of 857 Chinese model releases: only 3.6% ever published a safety result.
- Probability without a novel
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
A candid field note on Claude Science, including the artefacts the agents missed and a human caught.
- False fronts, Category 5
The richest primary-source case study yet of AI-assisted influence operations.
- When the drop is process, not a model
A dense weekly map of a week defined by math standards and safety firings; follow the receipts.
Today’s announcements mostly consolidate. Anthropic routes Claude into critical infrastructure through incumbent contractors, Google gives its agent a badge inside the corporate graph, OpenAI shows how much it can see of ChatGPT’s covert users, and Arena becomes the paid scoreboard all of them share. The dispersal is at the edges: Manus refinanced in China, copies of Jev, and research harnesses that anyone can run. Distribution and oversight are concentrating; capability still leaks outward.