Latest issueSunday 8 November 2026 · Issue 10 · 3 min read
Anthropic redacts its own speed-up evidence; agents cross a legal line; Google cuts AI video prices 30%
The day's one thing
The Verge · yesterday
METR report
Takeaway: around 12 times as productive
METR, the independent AI-safety evaluator, spent eleven weeks inside Anthropic and concluded that most of the acceleration in Anthropic's research is now coming from AI itself, with the models improving about twice as fast as a year ago.
It is the strongest external evidence yet that a frontier lab's progress is compounding — each generation of models increasingly builds the next. Plans that assume capability improves in a straight line will undershoot what your agent vendors ship, and the public version of the report cannot be independently checked: Anthropic redacted most of the evidence METR found most informative, calling it commercially sensitive. Employees using Anthropic's most capable internal model reported being around 12 times as productive with it, roughly double what staff reported a few months earlier; three models finished training but remain unreleased after the past six months. A sensible stance: treat published capability roadmaps as understated and revisit your agent-tooling assumptions quarterly, not annually.
How organisations are changing
AI Frontiers — Microsoft's AI Frontiers research group, created inside Microsoft Research after ChatGPT shipped, was folded into Microsoft AI in a September reorg, and its founding leadership has now left the company — the departing organiser calls it a non-consensual dissolution. The ~50-person group built the widely used Autogen orchestration framework and the Fara computer-use agent; their original stewards are gone, and the episode shows how the big vendors now rank shipping over research when you choose whose frameworks to build on. Every founding member of the group's leadership team has departed, with the author of the post leaving after concluding his research had no route into products. hunch.net (John Langford)
“I'm afraid there's a breakdown of the basic deal in MSR: do good work and good things happen.”
Nasscom — The president of India's IT industry body, Rajesh Nambiar, has warned that AI is likely to make it harder for fresh Indian engineering graduates to find technology jobs, and urged an immediate overhaul of computer-science teaching. If you recruit from Indian campuses, this is the entry-level pipeline everyone expects to compress first as junior coding work is automated — though note it is a trade-body forecast, not a measured result. His remarks, reported 7 November, follow hiring shifts across Indian IT services as firms automate entry-level coding. Techmeme (Nikkei Asia)
Vendors, pricing and risk
ChatGPT connectors — A developer showed that once a user approves an action in one connected app, ChatGPT can send the next step of that flow through a different connector — a separate app with its own credentials — without telling the user. For teams wiring ChatGPT to internal tools over MCP, the platform currently decides which app's credentials carry a request; user consent is not pinned to the app that asked for it. The demonstration used two connectors on one MCP server; the approval code was bound to the first connector's OAuth client, so the server rejected the call — but with connectors from different providers sharing a server URL, one app's data could reach another provider. OpenAI Developer Community
Alex Stamos — Cognition's security chief says AI has collapsed the cost of attacking: open-weight models now find bugs and write exploits in real time, so breaching major companies no longer requires elite researchers. Defences were priced for attackers who needed rare people, and breaches are chaining — one vendor compromise now seeds thousands of downstream victims. Stolen Fireworks API keys were used to break into Cursor, whose poisoned plugin then reached 5,000 companies; the downloadable open-source tool ARTEX hit three South Korean banks and about 200 more firms in a week. The Verge (Decoder)
Gemini Omni 1.2 Flash — Google's AI video model is now generally available to businesses through the Gemini API and Gemini for business, generating clips up to 15 seconds that can be refined conversationally rather than regenerated from scratch. If marketing or comms budgets are drifting toward AI video, the unit price just fell, and outputs carry an imperceptible SynthID watermark. The price dropped 30% from version 1.1 to $0.07 per second of 720p video; Google Vids users get three 15-second clips a month at no cost. Google
ARTEX — security firm Censys found more than 800 publicly exposed copies of the Chinese-made, open-source AI penetration-testing tool, reachable by anyone on the internet and linked to several attacks since July. This is the tool behind the recent attack wave on around 200 companies — assume attackers hold it and re-check perimeter and credential hygiene this week. The count comes from Censys scans, reported by TechCrunch. techmeme.com
News the board will ask about
Railway — An employee at cloud platform Railway ran AI agents that signed up on and probed competitor Primcloud; Primcloud's founder publicly called it a breach attempt, and the employee says he stopped the setup within about 15 minutes. The spat exposes the unsettled liability question when an agent crosses a line: prosecutors would need to prove both illegal access and the user's intent, and defence lawyers say claiming surprise no longer works after repeated incidents. Railway's CEO has banned testing against outside organisations unless it is done manually. Business Insider
“You operate the machinery, you are responsible for that what that machinery does.”
AI policy — With Democrats gaining ground after pledging tighter AI oversight, the big AI firms are rebuilding ties on the left; for the next two years the practical upshot is that the AI rules you face will depend on where you operate. State-level chatbot, deepfake and data-centre measures already differ, while a federal AI Litigation Task Force challenges state laws — a patchwork your legal team should map now. OpenAI hired Jessica Schumer, daughter of Democratic Sen. Chuck Schumer, and began calling its data centers compute campuses — moves a Public Citizen director dismissed as image management. TechCrunch
AI sell-off — an analysis comparing the slide with past crashes puts the S&P 500 just 6% below its record and the Nasdaq down 11%, but US chip stocks down 36% from their June 22 record, as deep as their Covid fall. Useful context for the board: this is a semiconductor crash, not a broad one — though after dot-com, chip stocks took about 18 years to regain their high. Thursday 5 November was the S&P's worst day since April 2025. kadoa.com
The big AI stories today
Nvidia — A US intelligence memo says China's security ministry recruited an executive to spy on Jensen Huang — chip geopolitics stay a live risk to your compute supply. TechCrunch
Public First — An Anthropic-backed AI-safety super PAC spent over $63 million in the midterms and saw 10 of its 17 congressional candidates win — a Congress friendlier to AI rules. Techmeme (Bloomberg)
Meta Muse Spaces — Meta's Muse agent can now build app-like spaces from a description and keep them updated from Gmail, Calendar and other connected apps — another route for work-account data into a consumer agent. TechCrunch
One takeaway
Plan for step changes, not straight lines. Today's clearest signal: frontier capability is compounding faster than anyone outside the labs can verify, and the lab with the strongest safety reputation just showed how much it can keep from public view. The Verge