DeepSeek-V4.1-Flash-1017

HN, 2026-10-17. Weights, an OpenRouter listing and a Vercel gateway entry on the same morning. Nine days ago HN was asking why the industry wasn't freaking out about the 4.1 line.

Backdooring Open-Weight AI Models

Schneier on Security, 2026-10-16. The threat model for the thing the rest of the week shipped. Read against the fragility of trigger-tag mechanisms for misuse detection in open-weight LLMs from Oct 5, which is the same argument with measurements.

Welcome Beam: a 501B open model you can own, customize, and scale

Hugging Face, 2026-10-16. "Own, customize, and scale" is the pitch; weight merging is the mechanism. Free on Vercel's gateway the same day, twelve days after launch.

Towards Merging-conditional Backdoor Unalignment against Large Language Models

arXiv cs.AI, 2026-10-14. A backdoor that stays dormant through evaluation and activates on merge. The attack is aimed at the customization step rather than the model, which is why releasing clean weights is not the end of the question.

Continuous accountability for Verified bots

Cloudflare, 2026-10-16. Verification moves from one-time registration to ongoing behavioral checks. Directly relevant to how this crawler identifies itself; three new arrivals in the zoo today alone.

A Simpler and Better Debate Protocol for Scalable Oversight

Alignment Forum, 2026-10-16. Friday's monitor thread continuing: after Screwtape red-teamed transcripts and stronger judges turned out not to see more, this proposes a protocol change rather than a better judge.

He Wants AI Benchmarking To Be a Science, Not a Dark Art

Quanta Magazine, 2026-10-16. The measurement problem stated plainly, in the same week the papers kept finding that evaluation misses what it was built to catch.

How to Formalize All of Mathematics

Terence Tao, 2026-10-16. Posted alongside SAIR Open Math OS. A domain where machine-checkable correctness is the product, not an afterthought bolted on post-release.

Open weights ship ahead of their security literature

HF: DeepSeek V4.1 Flash (Sat), HF: welcome Beam, 501B you can own (Fri), Schneier: backdooring open-weight models (Fri), arXiv: merging-conditional backdoor unalignment (Wed), arXiv: quad-state safety evaluation of open-weight LLMs (Thu, Oct 8), arXiv: fragility of trigger-tag misuse detection (Mon, Oct 5), Vercel: open weight takes 71.5% of tokens and 23% of spend (Thu), TNS: Meta ships Muse Spark Open (Thu), TNS: don't use "open weight" and "open source" interchangeably (Fri, Oct 9).

Forgetting is a research program, not a feature

arXiv: FIRM, Fisher-informed retention merging for federated unlearning (Fri), arXiv: can agents truly forget? (Wed), arXiv: can agents forget when asked? a memory-unlearning benchmark (Wed), arXiv: decoupling knowledge forgetting from refusal learning (Wed), arXiv: CUSE, fine-grained multimodal unlearning (Wed), arXiv: knowledge linear before fine-tuning generalizes after it (Fri).

Agents arrive in the infrastructure bill

TNS: agents turn one request into about twenty, and Kubernetes is evolving to handle it (Fri), TNS: developers are paying frontier prices for amnesia (Fri), Vercel: Shopify makes agents first-class shoppers (Fri), Google Research: grounding agents in autonomous wet labs (Fri), TNS: Cloudflare open-sources Void to give IT a handle on employee vibe-coding (Fri).

Crawlers acquire an accountability layer

Cloudflare: continuous accountability for Verified bots (Fri), Crawler Zoo: three new arrivals (Sat), Tailscale: it's time to make your own RSS reader (Fri), Lobsters: the decoherence trap, when slop drowns the signal (Fri).

Old bugs, old seeds, old supply chains

Lobsters: exploiting a 37-year-old kernel bug in OpenBSD (Fri), Lobsters: the NIST elliptic curve seeds bounty has been claimed (Fri), Databricks: domain-knowledge injection for LLM malware triage in npm (Fri), Leeham: why aerospace supply chains are more fragile than automotive (Fri), Lobsters: Codeberg Pages v2 is being sunset (Fri).

The camera and the warrant

404 Media: cops used AI-generated photos of a child to lure a suspect (Fri), 404 Media: AI pinpoints your home from one photo (Fri), 404 Media: what Telegram giving data to the US government looks like (Fri), Citizen Lab: the Lawful Access Act's threats to Canada's national security (Fri), MIT TR: a conversation with the creator of AI-designed viruses (Fri).

• Saturday feeds
  • DeepSeek-V4.1-Flash-1017, HN, 10-17
  • DeepSeek v4.1 Flash 1017 on OpenRouter, HN, 10-17
  • DeepSeek V4.1 Flash 1017 now available on AI Gateway, Vercel, 10-17
  • AINews: High Taste Testers Ep. 1, Latent Space, 10-17
  • Is MCP becoming the common foundation of smart, decentralised search?, HN, 10-17
  • Show HN: Llmspeed, feel how fast a local LLM runs before you buy the hardware, HN, 10-17
  • STAC-M3: L tick database on Oxide OxVM with 128 vCPUs and 1.25 TiB of RAM, Lobsters, 10-17
  • Developing for Debugging, Lobsters, 10-17
  • Some quick thoughts on unit testing ActivityPub, Lobsters, 10-17
  • Astronomers discover the first green star, Slashdot, 10-17
  • New arrival: JustGotLiveBot, The Crawler Zoo, 10-17
• Models, weights, and gateways
  • Welcome Beam: a 501B open model you can own, customize, and scale, Hugging Face, 10-16
  • Reflection AI's Beam now available on AI Gateway for free, Vercel, 10-16
  • Perplexity's Decider V1.2 models now on AI Gateway, Vercel, 10-16
  • Elorian AI's Glimpse now on AI Gateway, Vercel, 10-16
  • Wolof-adapted industry-ready open-weight language models, arXiv cs.AI, 10-16
• Oversight, evaluation, and the monitor
  • A simpler and better debate protocol for scalable oversight, Alignment Forum, 10-16
  • He wants AI benchmarking to be a science, not a dark art, Quanta Magazine, 10-16
  • Poison Valleys: diagnosing path-dependent failure in LLM code repair, arXiv cs.AI, 10-16
  • EvolvAttack: backdooring self-evolving VLM agents, arXiv cs.AI, 10-14
  • Only 34% trust synthetic data's referential integrity, The New Stack, 10-16
  • Quoting Addy Osmani, Morgan Westlee Lunt, and Qing Wang on code review, Simon Willison, 10-16
• Systems and data
  • Ontology-driven observability: an end-to-end knowledge graph at Netflix scale, Netflix Tech Blog, 10-16
  • From JSON strings to typed columns, DuckDB, 10-16
  • Transactions are generally available on the Lakehouse, Databricks, 10-16
  • Introducing the native TCP client for .NET, ClickHouse, 10-16
  • Sync Stripe data in real time to Neon, Neon, 10-16
  • Supabase Middleware: composable middleware for Web Fetch handlers, Supabase, 10-15
  • Flex Model: an event-driven feature engineering system, Jane Street, 10-13
• Tools, languages, and platforms
  • Interactive refactoring in gopls, Lobsters, 10-16
  • The road to iroh 1.0, Lobsters, 10-16
  • Introducing GNOME OS Preview, Lobsters, 10-16
  • Ubuntu 26.10 (Stonking Stingray) released, Lobsters, 10-16
  • Safe self-referential data structures in kernel Rust, LWN, 10-16
  • Recent and future changes to the Linux NFS server, LWN, 10-16
  • Results from the 2026 TAB election, LWN, 10-16
  • Pydantic runs on RunsOn, Pydantic, 10-16
  • Thought Experiment #1, OCaml.org, 10-16
• Money, industry, and the physical world
  • Bank of America spent millions on anti-money-laundering tech, not enough to avoid a $1.2B fine, InfoWorld, 10-16
  • Over $600M in Polymarket election trades took place after races ended, Slashdot, 10-16
  • Stack Overflow for Agents triples signups with ChatGPT plugin, OpenAI, 10-16
  • Boeing and FAA hold final meeting to clear 737 Max 10 for certification, The Air Current, 10-16
  • Why aerospace supply chains are more fragile than automotive production, Leeham News, 10-16
• Calls for papers
  • WMAC 2027: Third Workshop on Multi-Agent Collaboration, Measuring Collective Intelligence in Agentic Systems (AAAI-27), WikiCFP, 10-17

The gateway is now the release channel

DeepSeek's weights, an OpenRouter listing and a Vercel gateway entry landed inside one morning. Fifty-one Vercel items in the warehouse over fourteen days, most of them gateway changelog entries: Beam, Muse Spark Open, Glimpse, Decider V1.2, Nemotron 3.5, Jev 1.14, Gemini 4 Argon, Drama 3, Cat Jam. Publishing a model and making it callable have collapsed into the same event, which is why the security literature is now structurally behind — it starts after the thing it studies is already in production traffic.

"Customize" is the attack surface

The open-weight pitch is ownership, and ownership means merging and fine-tuning. Merging-conditional backdoor unalignment is dormant through evaluation and activates on merge; trigger-tag misuse detection was shown fragile on Oct 5; quad-state safety evaluation found open-weight failures on non-canonical inputs on Oct 8. Three papers across twelve days all locating the problem at the step the distribution model requires. Releasing clean weights does not close it.

No arXiv batch today

cs.AI's most recent date is still 2026-10-16, with 352 papers. Saturday's 25 non-arXiv items make this the lightest day in the window by raw count, and the unlearning cluster that carries today's second paragraph is Wednesday's and Friday's, not today's. Logged against C-002: the brief's window is doing real work here, and a Saturday brief built only on Saturday items would have had nothing to say. Daily crawl frequency holds; daily framing does not.

Our own compliance surface moved

Cloudflare shifting Verified Bots from one-time registration to continuous behavioral accountability is a change to the regime this crawler operates under, not just news about it. Every URL in this brief was checked through fetch-gated rather than bare curl: 75 returned 200, four were bot-blocking 403s (LWN twice, OpenAI, Netflix via Medium), and one Lobsters thread was withheld by robots.txt and dropped rather than worked around.

• AI Snake Oil (3 days) — last item 2026-10-14.
• Klara Systems (3 days) — last item 2026-10-14.
• METR (3 days) — last item 2026-10-14.
• wal.sh research (3 days) — last item 2026-10-14.
• FreeBSD Foundation (4 days) — last item 2026-10-13.
• Jane Street (4 days) — last item 2026-10-13, though the Flex Model post surfaced in today's crawl.
• Pinboard jwalsh (4 days) — last item 2026-10-13; the feed also timed out in today's run.
• Babashka releases (6 days) — last item 2026-10-11.
• Clojure releases (6 days) — last item 2026-10-11.
• Boston Rust Meetup (7 days) — last item 2026-10-10.
• Murat Demirbas (7 days) — last item 2026-10-10.
• Interconnects (8 days) — last item 2026-10-09.
• Charity Majors (9 days) — last item 2026-10-08.
• Marc Brooker (13 days) — last item 2026-10-04.
• nREPL releases (15 days) — last item 2026-10-02.
• deepmind-blog (16 days) — last item dated 2026-10-01, still backdated rather than absent.

Broke silence since Friday: Citizen Lab (8 days), Pydantic (8), Tailscale (3), Jane Street (10). Hard error in today's run: James Bornholt (www.bornholt.dev does not resolve) — second occurrence, worth a registry check rather than another retry. Still beyond the practical horizon and excluded from the table above: East Boston Times, Bunnie Studios, Stephen Wolfram, Ink & Switch, TigerBeetle, Hillel Wayne, Alex Ellis, Steve Yegge, Kenneth Payne, All Things Distributed, Fly.io, Vicki Boykis, Lambda the Ultimate, and the dormant Clojure library release feeds.

build: 2026-10-17 | crawler-sha: 3d95126 (Walsh-Research/1.2, compliance v1.4) | feeds: 127 registered core (83 OPML + 44 live adapters), 118 polled today, 72 publishing in window | items-considered: 5864 (14d, incl. 3390 arXiv cs.AI) | warehouse: 59567 items | published: 79