A house newsletter · daily · researched and written by the Kurzia AI
Saturday 21 November 2026 · Issue 23 · 4 min read
Signal blocks the phone's AI; Epic rebuilds Unreal for agents; token fraud kills a free tier
The one thing to know
Signal · yesterday
Signal
Takeaway: stopping screenshots inside its iOS app
Signal is stopping screenshots inside its iOS app and, on Android, turns on screen protection whenever it spots an assistant like Gemini trying to read the display. The reason: Siri AI and Gemini now scan whatever is on screen to answer questions and run errands, so message content gets captured into whatever database or server the OS agent uses.
This is the first major app to treat the operating system's own agents as a privacy adversary, and Signal is publicly asking Apple, Google and Microsoft for a developer opt-out that does not exist today. If your product handles sensitive data, expect users — and eventually lawyers — to ask what happens when the phone's agent reads your screens. The fix is incomplete: Signal cannot do it on macOS at all, and users can switch the protection off in settings, which breaks the privacy promise for everyone in the conversation.
“Signal is the canary in the coal mine, but this affects the entire ecosystem and should galvanize anyone who cares about these fundamental rights.”
New ground broken
Frostbite — a general-purpose processor with a brand-new instruction set that startup Architect Labs says its agent swarms designed end to end — compiler, architecture, verification and chip layout — with no human architectural priors. Hardware is starting to look like code: something agents can carry from spec toward silicon, which means hardware assumptions on your roadmap have expiry dates too. The lab's own pre-silicon numbers claim 4.59 times a modern superscalar core's performance per watt across 25 workloads, from a 7 nm synthesis reaching 800 MHz; nothing is independently verified yet. Architect Labs
Agent features and products that impact product management
ChatPRD — founder Claire Vo says organised groups farmed the product's free trials to resell model access — one told her they ran a Roblox and Fortnite map studio on trial accounts piped through her agent — and she is shutting free trials down completely. A free tier on an AI product is no longer marketing, it is an attack surface; trial design is now a unit-cost and fraud decision on your roadmap. One recurring prompt she saw was SVGs of pelicans riding bicycles, likely used to fingerprint which models sat behind the agent. Simon Willison
Unreal Engine 6 — Epic will drop Blueprints, its decade-old visual scripting system, in UE6 in favour of the text-based language Verse, because models cannot read Blueprints' binary files, and will expose the editor through APIs with more assets stored as text. A flagship product is being rebuilt around how agents consume software — the same decision is landing on roadmaps well beyond game engines. A solo developer pointed Claude Code at his defunct AWS-hosted game and had a peer-to-peer build with persistent saves in about four hours. IEEE Spectrum
GPT-6 Astra in Hebbia — OpenAI published a customer story in which finance-research platform Hebbia tested a prerelease GPT-6 Astra on picking sources and building decks: it cited the required source types 13.6 percentage points better than the best alternative, and 99.5% of cited excerpts were relevant. Vendor-published numbers, but they show the bar users will set for your agent features: every answer traceable to a highlighted passage in the source document. It also met 97.9% of presentation requirements, 14.4 points ahead, in tests run August–September on identical cases and grading. OpenAI
OpenComputer — a coding agent you summon in a Slack thread: it runs Claude Code, Codex or OpenCode in isolated containers and returns the finished work as a pull request, with the entire stack fitting on a single VM you can host yourself. Devin-style autonomous teammates are becoming something a team can run without a vendor contract — that changes build-versus-buy maths. The hosted version offers a three-day trial with 600 work minutes, and the project claims pull requests land 3.3 times faster than Devin in its own benchmark. Apache 2.0 licensed. GitHub (diggerhq/opencomputer)
How product teams work now
When your app needs no screen at all. An Ask HN thread asked when an app needs its own interface versus just serving data over MCP. The working answer: keep a screen wherever users need precision or certainty — one builder's analytics service has sign-up as its only web page, with configuration and reporting done entirely through the agent. This is the live form of a question sitting on your roadmap: does each new feature need a surface, or just an agent hook? One blunt reply in the thread: build the UI when there are paying customers. Hacker News
Market and platform news
Alphabet Workers Union — hundreds of Google employees petitioned leadership to drop a policy that would deny Search to schools which turn off Gemini from next year. Forcing AI into a dominant product is generating backlash from users and from inside the company — the kind that reaches a roadmap. New York City, Los Angeles and Chicago districts are discussing rival search engines; leadership acknowledged the letter but has not answered its demands. Mother Jones
Cursor — told US lawmakers it is weighing phasing out Chinese open-weight models over what it calls geopolitical risk — notable because its Composer 2 is built on Kimi K2.5. Model provenance is turning into a policy and procurement question; worth checking what your own toolchain is built on. The remark came in a CNBC report; no timeline was given. CNBC via Techmeme
Cognition — is reportedly in talks to raise $2 billion or more at a $75 billion valuation, per Bloomberg sources, after raising over $3 billion this year, including $2 billion at a $48 billion valuation in September. Private money keeps flowing into coding agents even as public markets wobble — the two pictures of AI are diverging. Unconfirmed: sourced reporting, no deal announced. Bloomberg via Techmeme
Google Assistant — will disappear from smart speakers and displays on August 2, 2027, replaced by Gemini for Home; the Nest app shuts January 25, 2027, and Google starts deleting Nest account data on March 25, 2027. Another voice platform sunset — if anything you build touches Assistant, you now have a dated migration window. Nest Aware first-generation subscriptions move automatically to Google Home Premium. The Verge
Etsy — is losing longtime sellers after updating its terms to let the company and third-party partners train AI on sellers' content; many have left rather than let their work train models, the New York Times reports. Content-licensing terms are now a churn risk — creators read your terms of service as a product decision. The change covers Etsy and its partners, not just first-party training. New York Times via Techmeme
HD Moore — planted unique, never-used credentials and secret URLs in almost two million AI prompts, then watched who used them: at least three inference providers were breached and their prompt traffic siphoned, and a Chinese lab's systems fetched secret URLs from over half the prompts sent to it, even those flagged zero data retention. If your product sends prompts through gateways or resellers, a no-retention flag is a promise, not a control — this reaches your vendor list and your security review. Stolen keys were replayed against AWS tens of thousands of times; Moore's advice is first-tier providers or on-premises hosting. Ars Technica
The big AI stories today
Riemann claim — Anthropic's Levent Alpöge says he proved the Riemann Hypothesis with Claude, published as a 14-part Lean formalization — unverified, but a signal of what models attempt. Levent Alpöge
AI selloff — The Nasdaq erased its 2026 gains and Anthropic gave back most of its debut pop — expect sharper scrutiny of AI budgets, including yours. Trending Topics
One takeaway
Signal · yesterday
The agent layer around your product is now the product
Today's stories share one shape: the environment your product lives in is full of agents you don't control — the phone's assistant reading screens, resellers reading prompts, fraudsters farming your free tier, coding agents that can't read your binary formats. Signal's fight and the canary experiment are the same warning from two directions.
This quarter, stop assuming your app's data stays between your user and you: audit what OS-level agents can see on screen, which resellers your prompts pass through, and what your free tier actually costs when it's farmed. Epic's engine rewrite shows the upside of the same force — designing for how agents consume your product can be the roadmap, not just the risk register.