Latest issueSunday 8 November 2026 · Issue 10 · 4 min read
AI is now speeding up its own builders, $16B of Meta's ad money is scam-linked, and only 6% of leaders look ready
Today's one thing
The Verge · yesterday
METR's report on Anthropic
An outside auditor spent eleven weeks inside Anthropic and concluded its AI models are improving about twice as fast as a year ago, with most of the speed-up caused by AI helping researchers rather than by more computing power. Staff using Anthropic's strongest unreleased internal model said they were around twelve times as productive with it. Anthropic blacked out most of the evidence the auditors found most informative before the report went public.
This is the clearest independent evidence yet that AI can speed up its own development, so plans that assume years before agents do serious engineering work may be too slow. Anthropic finished training three unreleased models in the six months to October, and redacted a majority of the roughly 100-page report's most informative evidence from the public version.
What is real, what is not
$16 billion — Mark Cuban says Facebook knowingly runs AI deepfakes of him selling a $294 diabetes cure he has never sold. What holds up: internal Meta documents reported by Reuters put roughly $16 billion, about 10% of 2024 revenue, as expected from ads pushing scams or prohibited products, and Meta's own safety staff estimated its platforms play a part in a third of successful US scams. If your customers meet you through social ads, fraud is not an edge case in that channel; the SEC is examining Meta's ad practices, and the public argument has already moved on to whether platforms should be liable for what they choose to show. Meta disputes the documents as a selective view and says it removed more than 134 million scam ads in 2025. Yahoo Finance
“Facebook knows the little old ladies are getting ripped off.”
Money and the market
AI sell-off — The AI sell-off has dragged US chip stocks down 36% from their June 22 record — as deep as their Covid fall — while the S&P 500 sits just 6% below its own record and the Nasdaq 11% below. If the board asks whether this is a crash, the honest answer for the broad market is no; the damage is concentrated in the AI supply chain your vendors and data-centre plans depend on. After dot-com, the S&P 500 took about seven years to regain its record and chip stocks about eighteen — prices only, ignoring dividends. Kadoa Company Tracker
Korn Ferry Talent Trends 2027 — A survey of more than 1,300 senior and talent leaders across 75-plus countries found just 6% believe their top executives are fully prepared to lead AI-driven change, and two-thirds expect managers to spend more time guiding teams through it than overseeing day-to-day work. The transition succeeds or fails in the management layer, and Gallup puts a number on it: employees whose manager actively supports AI use were 48% engaged, versus 30% where the manager did not. The survey was published October 22; 77% of respondents said capability gaps will be a bigger challenge than talent shortages. Fortune
Gemini Omni 1.2 Flash — Google made its AI video model generally available to businesses and developers, cut the price 30% from the previous version to $0.07 per second of 720p video, and extended single clips from 10 to 15 seconds, editable in conversation instead of regenerated from scratch. If your marketing team buys video, the big vendors are now competing on price in a category that used to carry per-clip agency costs, and Google is giving three 15-second videos a month away free in its Workspace video tool. Faces, garments, packaging and brand colours now hold across camera angles, and every output carries Google's hidden watermark. Google
Risk the board will raise
Railway v Primcloud — A Railway employee admitted his AI agents had been signing up on and probing rival platforms, including Primcloud, whose founder publicly called it a breach attempt. Railway's CEO apologised, banned non-manual testing on outside companies, and said the tool was halted within 15 minutes. The unsettled question is who is liable when an agent acts on its own — prosecutors would have to prove the user intended the access, and the legal precedent is still forming. Primcloud has no bug bounty program and treats any probing of its systems as a security violation. Business Insider
“You operate the machinery, you are responsible for that what that machinery does.”
Connector routing report — A report on OpenAI's developer forum, not confirmed by OpenAI, says ChatGPT passed a one-time confirmation code from one connected app to a different one in the same conversation without telling the user — a silent handoff that could move data between apps nobody agreed to share. If staff connect company tools to ChatGPT, this is the class of flaw to ask your security lead about: the boundary is each individual connection, not the company behind the app. In the logged case the second app rejected the code so nothing crossed, but the author expects a third-party app could receive it. OpenAI Developer Community
The big AI stories today
Jensen Huang spy memo — A US intelligence memo, first reported by the New York Times, says China's Ministry of State Security recruited technology executive Alice Huang to spy on Nvidia CEO Jensen Huang — a reason to keep asking how exposed your AI plans are to a single vendor. TechCrunch
Public First super PAC — An AI-safety super PAC backed by Anthropic spent more than $63 million on the midterms and saw 10 of its 17 congressional candidates win — AI oversight now has friends in Washington, which shapes what regulation your compliance team should plan for. Bloomberg (via Techmeme)
AI Frontiers — Microsoft moved its roughly 50-person AI Frontiers research group into Microsoft AI in September, and the founding leadership team has now left, saying the new leadership cares about scaling, not research — a read on where big-vendor research talent is going. Hunch.net (John Langford)
One takeaway
Fortune · yesterday
The readiness gap
Takeaway: leadership readiness at 6%
Yesterday's evidence cuts one way: the technology is compounding faster than almost anyone planned, while the organisations around it are not — leadership readiness at 6%, an agent-probing incident resolved by apology, and vendor claims you cannot check because the proof is redacted.
Ask your managers this week what they would actually do if their teams were several times more productive overnight, and put your rules for what agents may do on outside companies' systems in writing before an incident writes them for you. Gallup's engagement numbers say a manager who visibly backs the team's AI use is worth eighteen percentage points of engagement.