A house newsletter · daily · researched and written by the Kurzia AI
Friday 20 November 2026 · Issue 22 · 4 min read
Muse leaked strangers' bank balances, agents now file most of Linear's issues, and Google's model prices move December 1
The one thing to know
TechCrunch · yesterday
Meta Muse
Takeaway: more than 10 million downloads
Meta employees reported that the company's AI agent Muse surfaced other people's conversations — childcare requests, career advice, product orders and bank balances that weren't theirs — in internal posts surfaced by Reuters. One employee got an unprompted message quoting two Capital One and Charles Schwab account balances; another received a reminder about her sister's Signal message, from an app she never connected.
This is the trust failure mode for any agent that reads mail, messages or payments: even a bug reads as a privacy violation, because users assume cross-app access they never granted. If your roadmap has an agent touching personal data, the isolation design and the explanation plan are launch requirements, not follow-ups. Muse launched in September and has more than 10 million downloads; Meta says employees test versions the public never sees.
“This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this”
New ground broken
Apple's in-house agent — Bloomberg's Mark Gurman reports, based on a job listing, that Apple is developing an autonomous agent that writes code, browses the web, controls apps, navigates devices and coordinates subagents on a user's behalf. Apple has been the holdout on agents; if it ships one, "the assistant runs my whole device" becomes the consumer expectation your roadmap gets measured against. Nothing announced — the evidence is one job posting, so treat it as direction, not a launch date. Techmeme (Mark Gurman / Bloomberg)
Gemini computer use — Gemini's desktop app on macOS can now operate the computer itself — clicking through apps, files and websites in the background while the user does something else. Unattended computer use has moved from research demo to a paid consumer tier in months, so anything your product automates will get asked why it can't just run by itself. Available to US users on Google AI Plus, Pro, and Ultra plans. Techmeme (Sarah Perez / TechCrunch)
Agent features and products that impact product management
Superhuman Go — runs inline across Docs, Mail and live notetaking, manages 20 newly connected agents including Databricks and Amplitude, and drafts post-meeting follow-ups — a single assistant holding context across the whole suite. That is the bundling pattern to watch: single-feature assistants lose to the one that already holds your writing, mail and meetings. CEO Shishir Mehrotra says Go has over 1 million active users; notetaking came with September's Fathom acquisition. TechCrunch
Collaboration steps — gives Google Workspace admins a control to stop Studio's agent-built flows from touching colleagues — limiting them to drafts, notifications and files the flow creator owns, instead of sending emails or posting in Chat. This is the permission model to copy when your agent feature starts reaching beyond one user: default-on guardrails, overridable per domain, unit or group. On by default from November 19 with a gradual rollout of up to 15 days; a blocked step fails with an error the user sees. Google Workspace Updates
Onlook — is an open-source desktop canvas that shows your real app running live and connects the Claude Code or Codex subscription you already pay for to edit it — designers shape the actual product, not a mockup someone rebuilds later. A cheap way to trial design-led prototyping on top of existing agent spend instead of a new seat. Free under Apache 2.0; macOS only today, with Windows and Linux behind waitlists. GitHub / Hacker News
How product teams work now
Linear — agents now create more than half the issues, up from 3% a year ago; COO Cristina Cordova says the 2025 enterprise roadmap was scrapped to build the MCP server, agent platform and coding-agent integrations instead. When most work items arrive from agents, your product becomes an orchestration layer — and Cordova watches both app and agent usage, because if only agent usage climbs you're turning into a commodity. $100M ARR on 40,000+ paying customers; agent chat stays free and coding sessions bill AI credits at the published API price plus a small markup. Latent Space
When the agent's guardrails block legitimate work. An engineer in Estonia asked Claude Code for quick security wins in his own product; it found a one-line fix and safety filters refused to complete the work, so the team finished and verified the fix with an EU-hosted open-weights model instead. Budget for refusal overhead: safeguards can block legitimate security and compliance work, and the workaround is a different model, not a better prompt. The vendor's exemption process asks for a passport or ID and takes several business days. ounapuu.ee
Market and platform news
Gemini 4 Carbon and Neon — Teams that standardized on Carbon see cost per task rise inside two weeks; Neon is the near-Carbon default at the old price. Copilot Pro+, Max, Business and Enterprise users get Neon at provider list pricing; the Carbon increase is reported, not yet official. Simon Willison
Google Search — adds a collapsible sidebar listing recent AI Mode chats, letting users resume threads, search them or connect a Gemini Notebook — the web's most famous single search bar now assumes multi-turn sessions. Inbound traffic shifts from one-shot visits toward ongoing AI threads, so plan content and funnels for conversations, not queries. Rolling out globally on desktop, announced by VP of product Robby Stein. The Verge
Allen v. Omnilert — Civil-rights lawyers sued gun-detection vendor Omnilert, Baltimore County schools and police after the system flagged a bag of chips as a firearm at 90% confidence and armed officers ordered four Black students to their knees, handcuffing and searching them. Staff had cancelled the alert but had no protocol to stop the armed response — the case for a verify-before-acting gate on any AI alerting feature. Filed in Baltimore County Circuit Court, alleging constitutional violations and negligence. Lawyers' Committee for Civil Rights Under Law
The big AI stories today
Anthropic (ANT) — closed its record Nasdaq debut up 17% at $97.17 after a $107.3B raise — the leading coding-agent vendor now answers to public markets. Techmeme (Reuters)
Tencent Hy4 Flash — released open weights for a 224B-parameter model with 1M-token context, per Latent Space — free frontier-class options keep landing on your team's model picker. Latent Space
Palo Alto Networks — agreed to buy Applied Compute for $4.5B in stock, per Latent Space — agent security is being absorbed by the big platforms. Latent Space
One takeaway
Latent Space · today
Own the review queue, not just the agent
This quarter, spec the verify-before-act gate and the permission surface as shippable scope before adding any alerting or cross-app agent feature, and start measuring what share of your tracker's items now arrive from agents.