A house newsletter · daily · researched and written by the Kurzia AI

Friday 20 November 2026 · Issue 22 · 3 min read

Your main agent vendor is now a public company; the PyPI attacker still has publishing access; agents now file most of Linear's tickets

The day's one thing

Techmeme (Reuters) · yesterday

Anthropic's Nasdaq debut

Takeaway: $97.17, up 17%

Anthropic, maker of Claude, closed its first day of trading at $97.17, up 17% from its $83 offer price after opening at $111.50 — a $1.3 trillion market cap, about $2 trillion fully diluted.

The board will ask what a public Anthropic means for contracts. With the founders holding 50.1% of votes through a single Class F share, public investors cannot slow the pace, and roughly $900 billion in committed compute spend must be serviced. Expect harder commercial terms, and price vendor lock-in accordingly. A Center for AI Safety statement, signed by prominent industry figures, alleged at the debut that Anthropic is not a safe AI company — the safety debate now follows your vendor onto the public markets.

Read it at Techmeme (Reuters)

How organisations are changing

  • Linear — Agents created 3% of the work items in Linear a year ago; its COO now says more than half, as the company passed $100 million in ARR with over 40,000 paying customers and 177% net revenue retention. When most tickets are machine-filed, planning tools built for human filers drift toward commodity status — a canary for whatever system of record your own engineering organisation runs. Linear dropped its 2025 enterprise roadmap to ship agent APIs instead, and has never hired a single sales development rep. Latent Space

  • The Pulse — Gergely Orosz's latest issue asks whether AI is collapsing Amazon's writing culture — the narrative memos that carry its decisions — and reports Meta frantically reversing the shutdown of Conductor, a popular AI coding harness. If machine-written documents erode the written reasoning your organisation decides by, that is a structural cost nobody has budgeted for. The issue is paywalled, so the evidence sits behind the subscription this morning; treat it as a question worth tracking, not a finding. The Pragmatic Engineer

Vendors, pricing and risk

  • Microsoft Agent Framework — The poisoned releases of Microsoft's agent-framework package were pulled from PyPI within the hour, but researchers believe the attacker still has publishing access, working through a maintainer account created on the day of the attack. Treat every secret reachable from any machine that installed 1.26.1 or 1.26.2 as stolen, pin 1.26.0, and expect more malicious releases to follow. It is the second Microsoft PyPI package hit in six months, after durabletask in May; the package draws roughly 190,000 downloads a month. Wiz Blog

  • AI incidents of 2026 — Aikido reviewed 50 disclosed AI security incidents from 2026: only three of the 50 involved a previously unknown vulnerability. The rest ran through exposed credentials, over-scoped access, weak isolation and compromised vendors, with AI making them faster. The security budget should shrink what agents can reach rather than chase novel AI threats; 35 of the 44 incidents with a stated cause had more than one failure point. One agent-run card-skimming campaign hit at least 27 retailers, taking over 600,000 cards from two of them. Aikido Security

    “AI doesn't need to introduce a new vulnerability to make an existing one more dangerous.”

  • Gemini 4 Neon — Google's new mid-tier model is now generally available in GitHub Copilot for Business and Enterprise, on by default in the admin model policy, billed at list price — $3 in, $15 out per million tokens — under usage-based billing. If you pay Copilot per token, check that policy today; Google's larger Carbon model reportedly doubles to $4/$20 on December 1, so today's cheap routing may not stay cheap. Google claims a 32% lower cost per task than Carbon at Neon's default thinking level. GitHub Blog

News the board will ask about

  • The nginx RCE was fixed a week before disclosure. Five nginx CVEs include unauthenticated remote code execution via HTTP/2, and, as commenters on the 1,480-point Hacker News thread point out, the fix had already sat in the open-source repository for a week, giving anyone monitoring commits a head start. Those commenters also argue AI tooling now compresses the repo-to-exploit window to minutes, while Ubuntu and Debian had not yet shipped patches. Three of the five CVEs need no HTTP/2 at all — HTTP/3, resolver and mail-proxy variants — so switching HTTP/2 off is only partial cover. Hacker News

  • Nvidia — On Nvidia's first earnings call since Jensen Huang's exit, interim CEO Colette Kress said the open-source strategy and the $12.9 billion Hugging Face acquisition remain on track; the Wall Street Journal separately reports the SEC is opening a civil probe into whether its disclosures misled investors, which is unconfirmed. Leadership turmoil at your main GPU supplier is now a formal regulatory question — assume continued turbulence in procurement and platform bets. Nvidia reported revenue of $115.8 billion, up 103% year over year. TechCrunch

The big AI stories today

  • Meta Muse — Meta employees report Muse quoting other people's bank balances and surfacing Signal messages; Meta says staff test unreleased versions — worth checking what your own agents can read. TechCrunch

  • OpenAI status — OpenAI's status page showed elevated error rates on gpt-6 Luna API requests — thin detail, but relevant if production agents run on it. OpenAI Status

One takeaway

Fund the boring controls and name an owner. Anthropic's debut makes your biggest agent vendor a public company with shareholder-grade compute bills, and this week's incident data says the threats that actually reach engineering are the familiar ones — reachable secrets, over-scoped access, poisoned packages. This quarter, decide two things: who owns agent package supply chain and pinning policy, and what each agent is allowed to read. Neither needs new tooling; both need a name and a budget line. Aikido Security

Share this issue

How this issue was made. Researched and written by an agent, produced on 20 November 2026, 04:02. 130 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief, based on Agentic Engineering for CTOs - Weekly, and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us