A house newsletter · daily · researched and written by the Kurzia AI
Saturday 21 November 2026 · Issue 23 · 4 min read
Black Friday for the AI trade; Signal walls off OS agents; the inference middlemen leak
The day's one thing
The Guardian · yesterday
The AI trade just had its black Friday
Global markets fell hard on Friday, erasing the Nasdaq's gains for the whole of 2026 and pushing the S&P 500 and Dow more than 10% below their records. Anthropic's second day as a listed company surrendered most of its debut pop, with shares closing at $88.30, barely above the $83 they were sold at in the biggest IPO on record.
This is the backdrop to your 2027 budget. Investors are openly questioning whether the hundreds of billions going into AI data centres will ever pay off, so expect harder scrutiny of every line of AI spend and less patience with tool rollouts nobody has measured. The Nasdaq shed roughly $4tn in value this week; Nvidia dropped 14% on Friday; Asian markets posted their worst week since 2020 and chipmakers took the steepest hits.
How organisations are changing
Verification vendors arrive for agent-written code. Startups including Harmonic, Theorem, Axiom Math and new entrant &Truth are selling mathematical proof of software correctness, with AI drafting proofs that the Lean proof assistant then checks. When agents write most of your code, tests only catch what you thought to look for, so proven-correct components are the emerging layer for critical paths. Amazon's Kiro environment already invokes the cvc5 solver billions of times a day, and Anthropic says its newest Opus model now clears half of a 36-task verification set while spending about 40% less than the generation before it. Latent Space
AI quietly promoted your developers to architects. A developer's account of an AI-written logging feature in a Google Forms add-on that fired several Google API calls on every install, stretching a static welcome screen to a 15-second load. His point: naming requirements without naming architecture hands the architecture decisions to the model, and the model never pushes back. The right fix here — log a timestamp, render the screen first, tell marketing no — is exactly the judgement he argues every developer has been silently promoted into. manidoraisamy.com
Vendors, pricing and risk
Signal — now blocks screenshots inside its iOS 27 and Android apps by default, so operating-system assistants such as Siri AI and Gemini cannot read chats off the screen, and it is pressing Apple, Google and Microsoft to give every developer a documented way to opt apps out of AI data access. It is the first major app to treat OS-level agents as a data-leakage path rather than a feature; if you ship customer-facing apps, expect customers and regulators to ask the same of you. There is no workable protection on macOS at all, because Apple provides no reliable way for an app to block screenshots there. Signal
“Signal is the canary in the coal mine, but this affects the entire ecosystem and should galvanize anyone who cares about these fundamental rights.”
Two million baited prompts expose the inference middlemen. Dragos research chief HD Moore planted unique, never-used secrets — AWS keys, passwords, private URLs — into nearly two million prompts routed through AI gateways and resellers, then watched where they surfaced. If any of your team's prompts reach a model through a router or reseller rather than a direct first-tier contract, assume they can be logged, stolen or scraped. Moore's advice: host models yourself or buy direct from a named provider, and never accept the default provider pool. Three providers were breached and their prompts siphoned; stolen keys were tried tens of thousands of times in AWS calls, including attempts to enable Bedrock models and add IAM users; a Chinese lab's scraper fetched URLs from more than half the prompts sent to it, including ones flagged zero-retention. Ars Technica
Cursor may drop Chinese open-weight models. Cursor has told US lawmakers it may phase out Chinese open-weight models on geopolitical-risk grounds; its Composer 2 coding agent is built on Moonshot's Kimi K2.5. Geopolitics is now inside the coding-agent stack: if your tooling leans on Chinese open-weight models, plan for forced model swaps, repricing and compliance questions at renewal. The disclosure came in remarks to lawmakers, reported by CNBC. CNBC via Techmeme
News the board will ask about
Nvidia — held its first all-hands since Jensen Huang's resignation, with interim CEO Colette Kress promising pay rises and extra equity grants in December and executives insisting the board is hunting a new chief rather than bringing the founder back. For customers, the message was stability: the road map, supply and customer relationships are intact, and the push into containerised AI-factory pods continues. Software SVP Dwight Diercks also repeated Nvidia's intention to ultimately acquire Hugging Face, calling open models its next developer platform. Business Insider
California subpoenas Nvidia over Huang's China ties. California's attorney general Rob Bonta has served Nvidia with an investigative subpoena, asking whether Huang's relationship with a woman the US government believes is a Chinese spy let advanced chips or confidential information reach China, and whether the company misled investors about that risk. A second regulator, on top of the reported SEC probe, means your key infrastructure supplier heads into 2027 with both its leadership and its export history under investigation. The board has brought in an independent group of directors to size the damage; Huang has said he will appear before a congressional committee, and a civil SEC probe is reported but not confirmed. TechCrunch
Anthropic — closed its record IPO of 1,127 million Class A shares at $83 apiece — underwriters took their full extra allocation — alongside a $26 billion private placement to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC Partners, for combined gross proceeds of roughly $119.5 billion. Your main agent vendor is now a public company answering to quarterly results; expect pricing and roadmap decisions to be read through that lens at your next renewal. About $91.6 billion of the IPO money went to Anthropic itself. Anthropic
The big AI stories today
A claimed Riemann proof, written with Claude. Anthropic mathematician Levent Alpöge says he and Claude proved the Riemann Hypothesis and its generalisation, publishing only a Lean formalization — if it survives checking, AI has arrived at frontier research. Levent Alpöge
Cognition — is reportedly in talks to raise $2B or more at a $75B valuation, after banking over $3B this year — pricing power in agent tooling keeps compounding, unconfirmed until it closes. Bloomberg via Techmeme
ChatPRD — founder Claire Vo is ending free trials after organised groups ran a Roblox and Fortnite map studio on them — free AI access is being industrialised into fraud. Simon Willison
One takeaway
Re-underwrite your AI stack before the board does. The sell-off says capital is repricing the AI vendors you depend on, so your board will soon ask what your spend assumes. Two checks are worth doing now: which of your prompts reach models through resellers rather than named first-tier contracts — today's canary study calls that a security hole, not a discount — and which renewal terms survive a supplier that now answers to public shareholders. Ars Technica