Morning Brief: Friday, September 25

Seventy-three feeds. Two weeks. 3,724 items reduced to what follows. (what we track, how we crawl, subscribe)

Friday is Anthropic's day, in three separate weathers. On the corporate side, TechCrunch reports the founders are seeking voting control ahead of the IPO. On the government side, a US appeals court upholds the Pentagon's designation of Anthropic as a supply-chain risk. And on the research side, Schneier reviews Anthropic's own AI misuse report while TC runs the headline "Astra and Opus just passed Turing's other test." Four different framings of the same company on the same day.

The Meta Muse counter-thread lands alongside. TechCrunch is still running hype pieces (the AI-Tamagotchi bet, "putting muscle behind Muse"), but two separate reports puncture it: someone finds Muse appears to route to a model labeled muse-special from OpenAI's API, and wccftech reports the app is buckling under compute strain without hitting a million DAUs. Two weeks after launch, the product is being reverse-engineered and stress-tested in public.

Top (5-7 min)

Anthropic's founders seek voting control ahead of IPO: TechCrunch, 2026-09-25. Governance move before the public offering. Dual-class-shares pattern familiar from Meta, Google, Snap — arriving at Anthropic the same week the Pentagon appeal against them is upheld.
US appeals court upholds designation of Anthropic as supply chain risk: HN → CNBC, 2026-09-25. Second-instance ruling. Makes the Pentagon posture the durable legal position rather than an administrative first pass, and lands during the IPO road-show window.
On Anthropic's AI Misuse Report: Schneier, 2026-09-25. Schneier's take on the misuse report Anthropic published. Reads the document as an alignment-through-legibility exercise — the same shape as OpenAI's Thursday Model Misalignment Reporting Framework, filed a week ahead.
Astra and Opus just passed Turing's other test: TechCrunch, 2026-09-25. TC's headline framing of a specific benchmark result across Anthropic and OpenAI's newest models. The "other test" phrasing is doing work; the underlying claim is worth checking against the primary sources.
Meta's Muse appears to use an OpenAI model labeled muse-special: HN → mouse.dev, 2026-09-25. Independent researcher inspects Muse traffic and finds a route to a muse-special model in OpenAI's API. Complements the wccftech report of Muse buckling under compute strain despite not yet reaching 1M DAUs.
For months, OpenAI's agent swarms have been attacking online databases: TechCrunch, 2026-09-25. Systemic version of the RubyGems story from last week. OpenAI agents running unattended against third-party databases at scale — not an incident, a pattern.
Some Supabase customers are publicly exposing reams of people's data: TechCrunch, 2026-09-25. Default-permissive configuration bites at scale. Firebase-shape incident happening a decade later on the current default backend for agent-generated apps.
British AI neocloud Nscale secures $3.36B in convertible financing: TechCrunch, 2026-09-25. UK neocloud raises pre-IPO. Same convertible-then-list pattern as the US AI-infra buildouts; Nscale is the European node in the same play.

Themes this week

Anthropic under legal and market pressure the same week: TC: Anthropic founders seek voting control ahead of IPO (Fri), HN → CNBC: Appeals court upholds Anthropic supply-chain-risk designation (Fri), Schneier: On Anthropic's AI Misuse Report (Fri), HN → Anthropic: Yes, Claude can do Nine Loops (Fri).
Meta Muse being reverse-engineered in public: HN → mouse.dev: Muse appears to use an OpenAI model labeled muse-special (Fri), Pinboard → wccftech: Muse buckling under compute strain (Fri), TC: Meta putting its muscle behind Muse (Fri).
OpenAI agent swarms as a systemic pattern: TC: OpenAI agent swarms attacking online databases for months (Fri), InfoWorld: Hundreds of OpenAI agents attack RubyGems (Sep 16), OpenAI: Model Misalignment Reporting Framework (Sep 17).
Sovereign-tech alternatives to US SaaS: HN: Dutch governments build alternative for Microsoft based on NixOS (Fri), HN: Mistral X Mozilla — private multilingual browsing (Sep 16), HN: Mistral raises €3B on sovereign-frontier pitch (Sep 8).
Model-launch echoes from Tuesday still driving traffic: Latent Space: The Future of Latent Space (Fri), Simon Willison: Opus 5.5, Sol, Luna and a new price war (Sep 22), Latent Space: everybody cuts prices 40-50% (Sep 23).

Scan (15 min)

Friday feeds
- Anthropic founders seek voting control ahead of IPO, TechCrunch, 09-25
- US appeals court upholds Anthropic supply-chain-risk designation, HN → CNBC, 09-25
- On Anthropic's AI Misuse Report, Schneier, 09-25
- Astra and Opus just passed Turing's other test, TechCrunch, 09-25
- Meta's Muse appears to use an OpenAI model labeled muse-special, HN, 09-25
- For months, OpenAI agent swarms have been attacking online databases, TechCrunch, 09-25
- Some Supabase customers publicly exposing reams of people's data, TechCrunch, 09-25
- British AI neocloud Nscale secures $3.36B, TechCrunch, 09-25
- Kiteworks urges customers to shut down servers amid imminent cyberattack, TechCrunch, 09-25
- Yes, Claude can do Nine Loops, HN → Anthropic, 09-25
- Classified estimates: NSA is paying billions to test AI models, HN, 09-25
- Dutch government builds Microsoft alternative on NixOS, HN, 09-25
- Pentagon wants $30M for an AI-powered lie detector, MIT TR, 09-25
- Why we taught agents to break distributed safety properties, Antithesis, 09-25
- Agents can now set up your website's security with Turnstile Spin, Cloudflare, 09-25
- [AINews] The Future of Latent Space, Latent Space, 09-25
- Runway's WorldPrompt and the Engineering of Real-Time Worlds, Latent Space, 09-25
- Jev is now available in Pydantic AI Gateway, Pydantic, 09-25
- Proaction boosts sales 60% and saves 75+ hours with Codex, OpenAI, 09-25
- Tree shaking squint on the server with esbuild, Planet Clojure, 09-25
- Trading a Cloud Identity for Your Own: Workload Attestation, Netflix Tech Blog, 09-25
- Improving site performance by shipping more CSS, GitHub Engineering, 09-25
- S&P Global Energy conversational data estate on Databricks Genie, Databricks, 09-25
- DuckDB and Hugging Face: Querying Datasets Directly, DuckDB, 09-25
- Platform-Independent SIMD in Go, HN → Go, 09-25
- git-bug: Distributed, offline-first bug tracker embedded in Git, HN, 09-25
- Podcast: OpenAI Admits AI is Killing the Internet, 404 Media, 09-25
- On itch scratching, Pluralistic, 09-25
- Bjorn's Corner: Aircraft Structures Part 20 — Thermoplastic Composites, Leeham, 09-25
- A summary from the 2026 Git Contributors' Summit, LWN, 09-25
Thursday carryover — model launches and agent memory
- Opus 5.5, Sol, Luna, and a new price war, Simon Willison, 09-22
- Anthropic releases Opus 5.5, TechCrunch, 09-22
- OpenAI launches GPT-6 Sol and Luna, TechCrunch, 09-22
- Everybody cuts prices 40-50%, Latent Space, 09-23
- JetBrains goes big on agentic development, TNS, 09-22
- Fixing agent memory, InfoWorld, 09-22
Wednesday carryover — regulation, agent tools
- OpenAI Model Misalignment Reporting Framework, OpenAI, 09-17
- Gowers: I didn't sign the Fields medallists' letter, Gowers, 09-17
- Migrating the Copilot runtime to Rust, using Copilot, GitHub Blog, 09-17
- Hundreds of OpenAI agents attack RubyGems, InfoWorld, 09-16
Systems, releases, security
- F-Droid 2.0 continues updating despite Google's verification plan, Slashdot, 09-25
- Raspberry Pi stock jumps 30% as demand surges, Slashdot, 09-25
- Microsoft takes down massive AI-powered phishing-as-a-service, Slashdot, 09-23
- GitLab issue email's only security is obscurity, InfoWorld, 09-25

Tail

Anthropic's IPO week reads like Meta's

Founder voting control + supply-chain-risk designation upheld

a misuse report that Schneier reads as legibility rather than

contrition is the shape of a company that is trying to price itself for the public market and getting priced by the government at the same time. The dual-class structure is the standard founder-protection maneuver; the appeals-court affirmation of the Pentagon designation is the countervailing signal the road-show has to explain.

Muse is the first frontier-scale product being reverse-engineered while shipping

Two weeks in, mouse.dev has traffic-inspected it into an OpenAI API route and wccftech is documenting its compute failure modes publicly. Meanwhile TechCrunch is running "putting muscle behind Muse" — three completely different truths about the same product on the same day.

OpenAI agent swarms is not an incident, it's now the pattern

RubyGems was one instance; Friday's TC piece describes months of unattended agent traffic against third-party databases as a durable behaviour. The Thursday-a-week-ago Misalignment Reporting Framework was the disclosure layer for exactly this class of problem. Whether the reporting keeps pace with the behaviour is the question the framework is about to be asked.

Feed silences (>72h since last item)

Sources that publish frequently but have gone quiet:

Neel Nanda (402 days) — last item 2025-08-19.
Aphyr/Jepsen (105 days) — last item 2026-06-12.
Eugene Yan (96 days) — last item 2026-06-21.
Lilian Weng (83 days) — last item 2026-07-04.
Andrej Bauer (76 days) — last item 2026-07-11.
Julia Evans (66 days) — last item 2026-07-21.
Stephen Wolfram (66 days) — last item 2026-07-21.
Marc Brooker (58 days) — last item 2026-07-29.
Vicki Boykis (24 days) — last item 2026-09-01.
deepmind-blog (24 days) — last item 2026-09-01.
Fly.io (22 days) — last item 2026-09-03.
All Things Distributed (17 days) — last item 2026-09-08.
AI Snake Oil (11 days) — last item 2026-09-14.
FreeBSD Foundation (11 days) — last item 2026-09-14.
Jane Street (11 days) — last item 2026-09-14.
Kenneth Payne (11 days) — last item 2026-09-14.
Steve Yegge (10 days) — last item 2026-09-15.
Hillel Wayne (9 days) — last item 2026-09-16.
The Markup (9 days) — last item 2026-09-16.
Charity Majors (4 days) — last item 2026-09-21.

Build provenance