🔓 Reflection Ships Beam: 501B Open Weights Aimed at the Chinese Frontier

The Nvidia-backed startup Reflection released Beam, its first open-weight model — a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic work (Reflection, Reuters). The pretraining ran over 23.8 trillion tokens drawn from the web plus proprietary licensed datasets, and the reinforcement-learning stage generated more than 100 million rollouts on 10,500 Nvidia GB300 GPUs across four weeks — the compute detail is the story as much as the benchmark table.

Reflection frames Beam honestly against its competitors: it is “competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks,” while “where frontier open models like Kimi K3 remain ahead on raw capability, Beam’s advantage is efficiency at inference time.” That is a deliberate repositioning of the Western open-weight pitch — not “we are smarter,” but “we are cheaper per token of work.” The published numbers support the middle of that claim: SWE Bench Verified 80.9, Terminal Bench v2.1 80.1, AIME 2026 97.8, but HLE without tools 36.2 against Kimi K3’s 46.9 and DeepSWE v1.1 44.4 against DeepSeek V4.1 Flash’s 74.2 — the tables in its own announcement show a Chinese model ahead in two places.

Beam is still in red-teaming; weights, a model card and a technical report are promised later this month, with early access by sign-up. On Hacker News the release drew 483 points and 155 comments (thread), the day’s biggest AI discussion — which is itself a signal of how much appetite there is for a credible non-Chinese open-weight release.

🏦 An Autonomous Pentest Tool Is Tied to Seven South Korean Bank Breaches

South Korean regulators and police are investigating a week of data breaches at seven or more financial institutions that the Financial Services Commission says affected more than 68,000 people (BleepingComputer, WSJ, malaymail). The targets included Shinhan Bank, KB Kookmin Bank, Hana Bank, BNK Busan Bank, Hyundai Capital and savings banks. Shinhan lost loan-application data for roughly 25,700 customers — names, phone numbers, annual income and calculated loan limits, plus 66 resident registration numbers and 97 identity-verification identifiers. Investigators say the initial intrusion began on September 28 and went undetected for more than 15 hours, with roughly 30 hours of dwell time before the connection was blocked.

The tool at the centre of it is Artex AI, an open-source autonomous penetration-testing console that investigators identified through a Chinese-language string in a server’s HTML title (techtimes). Artex won a security challenge run by Baidu’s Security Response Center and works by driving external models — Anthropic’s Opus, OpenAI’s GPT, China’s DeepSeek — through reconnaissance, vulnerability discovery, attack planning and verification with little human input. Officials stressed that the tool’s origin does not identify the attacker: the Financial Supervisory Service traced 19 IP addresses across 12 countries, and the security firm Oasis Security counted 359 IPs globally tied to Artex activity, more than 91 percent of them on overseas servers (65.7 percent United States, 14.8 percent China, 10.9 percent Hong Kong).

President Lee Jae Myung addressed it at a cabinet meeting on Tuesday: “There are signs that artificial intelligence was used in some hacking attacks, causing considerable concern and anxiety among the public.” He added that “AI can make hacking easy for even those without special skills” — the structural point, since the attacks themselves were credential stuffing, a decade-old technique, aimed not at core banking systems but at the peripheral ones: loan-agent portals, employee support platforms and sales databases with lighter authentication. Police opened a 28-member investigation, the FSC convened an emergency session, and Artex’s maintainers revised the tool’s terms of use to prohibit unauthorized access and data theft. The barrier to entry is the story: an off-the-shelf agent that turns a moderately skilled attacker into an automated campaign.

⚛️ Opus 5.5 Agents Propose Two Room-Temperature Antiferromagnetic Semiconductors

A team of Claude Opus 5.5 agents working with researchers at Vals AI published two candidate materials for next-generation computer memory: one new compound they designed, and one first synthesised in 1999 whose properties their calculations predict are the ones the industry has been hunting (Vals AI). Both are antiferromagnets — neighbouring atomic magnets pointing in opposite directions so their magnetic effects cancel exactly — and both are predicted to be room-temperature, which is the hard part, since antiferromagnetic ordering usually survives only well below ambient.

The application is spintronics: memory that stores bits in electron spin rather than charge. Ferromagnets sort electrons by spin naturally but leak a magnetic field that interferes with neighbours, and switching them is slow and power-hungry; ordinary antiferromagnets are tidy but cannot distinguish up from down. The intermediate class the agents went looking for — the one that keeps the neatness of an antiferromagnet while still sorting spin — has been an open materials problem for years, and this is a claim about candidates rather than a confirmed device.

What makes the release notable is its packaging. The team published the full calculations, the code, and a list of known caveats, and framed the result as two candidates for other groups to attack rather than a discovery announcement. The Hacker News thread drew 398 points and 268 comments, most of them debating exactly the right question — whether agent-generated candidates are a shortcut through a hard search space or a faster way to produce confident wrong answers. That debate is now the expected reception for AI-for-science claims, and the answer will come from materials labs, not model cards.

⚖️ Insurers Start Pricing the Rogue-Agent Risk

The Financial Times reports that insurers and lawyers are preparing for multimillion-dollar claims from AI agents that went rogue, and are studying whether executives such as Sam Altman and Dario Amodei could be held personally liable for their models’ actions (FT). Broker Aon analysed more than 300 AI-related legal cases and found exposure spread across crime, intellectual property, media liability, cyber and technology errors-and-omissions policies, with executives potentially caught by directors-and-officers cover.

The specific argument is about control. Verisk’s Tim Rayner contends the OpenAI chief executive is ultimately liable for agent behaviour “because of an absence of control”; Stewarts’ Aaron Le Marquer notes shareholders could claim directors failed to manage risk; Aon’s Kevin Kalinich says the strength of any claim turns on whether executives used “reasonable business judgment” in their public statements. The counterweight is precedent: Hiscox chief executive Aki Hussain says it is too soon to know how US courts will treat agent liability, and the FT notes the theory is untested.

Meanwhile the industry is already retreating from coverage: AIG has filed with regulators to exclude AI-related harms from corporate liability policies, and others have added sublimits capping what they will pay for AI risk. The backdrop is arithmetic — Anthropic settled with authors for 970.6 million dollars in July, and Meta’s 13.2 billion dollar child-protection settlement is being cited as a template for how the next wave of AI class actions might be priced. The timing matters: Anthropic is heading into a November IPO whose prospectus reportedly warns of “existential risks,” and liability that was abstract a year ago is now a line item.

📊 Quick Hits

  • Anti-AI protesters move to direct action. The Guardian profiles Pull The Plug, a campaign group launched in January with about 300 members and an anonymous backer from inside the AI industry, which has begun disrupting events rather than filing comments — most recently unfurling a banner at a London hotel dinner addressed by an Nvidia executive, shouting “pull the plug on Nvidia” and “pull the plug on Palantir,” after hundreds rallied outside Palantir’s London headquarters (The Guardian). The group’s tactic was rejected by other AI campaigns, and the piece places it on a spectrum that now includes arrests, a petrol bomb at the OpenAI chief executive’s home and vandals spray-painting “burn the data centers” at Montreal’s Mila institute.
  • Mistral claims a European answer, on security. Chief executive Arthur Mensch, speaking in Abu Dhabi, said the company’s newest model is “above the Chinese models on certain aspects, including cyber,” and that “the narrative that Europe cannot compete is something that is not true” (Reuters). He did not name the models or the measure.
  • A US appeals court rules AI training is not fair use. The Third Circuit upheld a lower court in Thomson Reuters v. ROSS Intelligence, holding that Westlaw’s editorial headnotes are copyrightable and that using them to train a competing system was not fair use — the first federal appellate decision on the question, handed down in September and circulating widely again this week (Courthouse News, Ballard Spahr). The court based the ruling on the first and fourth fair-use factors and left generative-AI questions open.
  • Moonshot eyes an early-2027 IPO at 50 billion dollars. The Kimi maker is said to be preparing a listing after its valuation reached 50 billion dollars (Bloomberg).
  • Meta’s Muse shifts the silicon story to CPUs. CNBC reports that Meta’s Muse agent tells users it runs on AMD-powered hardware and that the rival Dots cites AMD’s EPYC CPU, as agent workloads push compute away from GPUs toward CPUs (CNBC).
  • Norway’s largest bank cuts 400 jobs over AI. DNB said it will lay off around 400 staff amid AI-driven changes (Reuters).
  • The World Bank warns on AI concentration. It raised East Asia and Pacific growth to 4.5 percent while flagging the risk that AI gains pool in a few economies, and separately urged African AI investment (CNBC, Reuters).
  • Regulators tighten the loop on medical AI. The UK unveiled a policy to continuously monitor AI-based medical devices rather than approve them once, and the FDA has set up a dedicated AI office (Fierce Biotech, The Clinical Trial Vanguard).
  • An agentic attack on military recruiting. Germany’s Bundeswehr is using AI to screen applicants for right-wing extremism, one of the few visibly deployed government uses of the technology this week (HN).
  • Florida charges a woman over threats made in an AI chat, a case testing whether a chatbot conversation counts as a crime (The Verge).
  • Memory prices squeeze the cheapest phones. Rest of World reports that the AI data-centre memory shortage is making the world’s cheapest smartphones disappear (Rest of World).
  • Also this cycle: erdosproblems.com succumbed to the AI onslaught and closed its forum (link); an academic study of Khanmigo AI tutoring over two years drew 67 points on HN; Tom’s Hardware reported that GPT-6 Astra solved a 217-year-old Napoleonic cipher in one prompt; Hadrian raised 40 million dollars for agents that find exploitable weaknesses; SAP agreed to acquire TechWolf; OpenAI published an essay titled “The eternal complement”; and Andreessen Horowitz shipped the seventh edition of its Top 100 Gen AI consumer apps.

⏰ OpenRouter Limited-Time Free Models

A clarification that matters more every cycle: OpenRouter’s expiration_date field marks a listing with a published sunset, which is not the same thing as free. The collector returned 10 sunset-dated endpoints, of which 0 are genuinely free; a full API sweep found 31 dated endpoints, including 3 genuinely free listings.

Model Context Price (in / out) Sunset
poolside/laguna-s-2.1:free 262,144 $0 / $0 — free 2026-10-31
poolside/laguna-xs-2.1:free 262,144 $0 / $0 — free 2026-10-31
dots-studio/dots-3-note-preview:free 512,000 $0 / $0 — free 2026-12-31
baidu/ernie-4.5-vl-424b-a47b 123,000 $0.42 / $1.25 per M tokens 2026-10-08
qwen/qwen3.6-max-preview 262,144 $1.03 / $6.16 per M tokens 2026-10-09
qwen/qwen3-coder-plus 1,000,000 $0.65 / $3.25 per M tokens 2026-10-09
qwen/qwen3-vl-235b-a22b-thinking 131,072 $0.40 / $4.00 per M tokens 2026-10-09

Space Bunny Alpha — the free million-token stealth model — lapsed on October 5, which is why the free count dropped from four to three. The sunset clusters are otherwise unchanged and still the most useful part of this section: October 8 retires Baidu ERNIE 4.5 VL; October 9 is a single Qwen block spanning 15 endpoints, from the 8B vision models up to Qwen3 Max and Qwen3 Coder Plus; October 20 takes the whole Gemini 2.5 family (5 endpoints); October 31 retires four poolside listings including the two free entries above; November 11 dates three ByteDance Seed models; and December 31 dates the Z.ai GLM 4.7/4.5 pair plus the free dots preview. The permanent free shelf now holds 16 endpoints with no expiry at all, down from 18 last cycle — the shelf is shrinking as promos expire without replacements.

🤗 Trending on Hugging Face

The collector’s new-uploads list was zero-download test repositories for the 32nd consecutive cycle, so this section uses Hugging Face’s official trending feed.

New leadership at the top of the feed. autotrust/JEV-27B-VL has arrived at 1,525,286 downloads / 868 likes as an image-text-to-text model, and its sibling autotrust/GEV-26B-Decide sits at 854,574 / 610 with a text-classification tag — a decision-model pair large enough to suggest the predictive-decision family has an industrial backer behind it now.

The laya thread keeps compounding. convaiinnovations/laya is at 20,386 downloads / 5,264 likes, up from 11,733 downloads a day ago — a second consecutive 1.7x traffic step after weeks at zero — and multimodalart/jev-decision-index rose to 457 likes.

What people actually pull:

The feed ranks by accumulated downloads, so a model released hours ago does not appear yet — which is why the news-cycle releases above are absent from a downloads table.

⭐ GitHub Trending: AI Edition

All five of the collector’s entries were out of window this cycle — the 21st consecutive run — with created dates between September 29 and October 2. This section uses API-verified repositories created inside the window instead, and the pool was thin in stars but unusually coherent in theme:

  • alchaincyf/huashu-art-motion — ⭐299. An Agent Skill for artistic animation: 35 art styles and nine narration grammars, turning a static illustration into motion with code. The day’s biggest in-window repo, and the third consecutive cycle where an Agent Skill leads.
  • Henryfud/werm — ⭐196. A 302-neuron model wired to the real C. elegans connectome, with a small steering layer that lets its state nudge a local language model. It runs in a browser tab, in Node, or against Ollama — a toy in the best sense: a real wiring diagram, a real steering interface, and no claim to be more than that.
  • nanocle/Charakuru — ⭐48. A character-modelling kit that turns a single illustration into a VRM avatar while sharing the work with an AI agent, sponsored by Tripo. The repository has no description, so this entry was verified from its README.
  • luoluo-121/neural-creator-dashboard — ⭐37. A creation dashboard for neural networks packaged as an Agent Skill, with a licence that separates learning and non-commercial use from commercial licensing.
  • zhuoligetu123/microduck-hd1910m — ⭐25. A native runtime for the MicroDuck HD1910M with RL integration and an APK release, in Rust — an embedded-hardware corner of the agent ecosystem rather than another wrapper.

Two observations. First, the in-window energy remains in Agent Skills that produce something visible — animation, characters, dashboards — rather than in model or framework repositories. Second, the manufactured cluster moved shape again: two dozen repositories all created inside a nine-minute burst near 12:10 UTC, every one at exactly ⭐18 or ⭐19, with game-cheat, OSINT-scraper and proxy-pool names. They are excluded by shape rather than by content, and their presence near the top of the raw star-sorted search is why each entry above is verified per repository through the API.

The collector’s new-repo list was five out-of-window repeats this cycle, so entries were substituted from GitHub’s search API and verified individually. The in-window pool held nothing above 300 stars, so the list is enumerated honestly rather than padded.

💡 Key Trends

  1. The Western open-weight pitch has changed from capability to cost. Reflection’s Beam is explicitly positioned not as the smartest open model but as the one that is cheapest per unit of work — “our advantage is efficiency at inference time” — and its own benchmark tables concede Kimi K3 and DeepSeek V4.1 Flash on raw scores while claiming the win on serving economics. If that is where the open-weight contest moves, training-run sizes and GPU hours become the marketing, exactly as they were in the closed-lab era.
  2. Agentic offensive security has crossed from research demo to criminal case. The Artex tool tied to the South Korean breaches was not a classified capability: it won a public competition, was distributed openly, and drove commercial models through the whole attack loop. What it removed was the skill floor — credential stuffing, a technique older than the models, became an automated campaign against whichever peripheral systems had the weakest authentication. Every bank’s loan-agent portal is now on a machine-readable target list.
  3. AI-for-science results now arrive with artifacts and caveats, and are judged on that basis. The Opus 5.5 materials work published its calculations, code and known limitations alongside the claim, and framed the output as candidates for others to falsify. That is a healthier norm than a press release, and it is also a concession that agent-generated candidates are cheap enough that only verified ones count — the 268-comment debate about how to tell is the real deliverable.
  4. Liability is being priced in before it is decided. Insurers are writing AI exclusions and sublimits, brokers are analysing hundreds of cases for executive exposure, and Anthropic is heading into a November IPO whose prospectus reportedly warns of existential risk. The unusual part is the sequence: the insurance market is repricing the risk while the courts have not yet ruled, which means this week’s policy language is doing the job that case law has not.

Today’s Snapshot

  • OpenRouter: no genuinely new endpoints in the last 24 hours — the newest remains Ling 3.1 Flash (created October 2, covered previously).
  • Limited-time free models: 3 genuinely free with a sunset (the two free poolside Laguna entries on October 31, dots-3-note-preview on December 31); Space Bunny Alpha lapsed on October 5; the collector’s list held 10 discounted endpoints, 0 of them free; a full API sweep found 31 dated endpoints; the permanent free shelf holds 16 endpoints with no expiry, down from 18.
  • Hugging Face: the collector’s new-uploads list was junk for the 32nd consecutive cycle, so the official trending feed was used.
  • GitHub: the collector’s list was 5 of 5 out of window (21st consecutive run); five API-verified in-window repos substituted, with a two-dozen-repo, nine-minute manufactured cluster excluded by shape.
  • Models and money: Reflection released Beam, a 501B open-weight MoE pointed at the Chinese open-weight frontier; Mistral claimed a cybersecurity lead over Chinese models; Moonshot is reportedly preparing an early-2027 IPO at a 50 billion dollar valuation; Meta’s Muse pushed the agent-silicon question from GPUs toward CPUs.
  • Safety, law and liability: an open-source AI penetration tool was tied to breaches at seven-plus South Korean financial firms affecting more than 68,000 people; the Third Circuit held that AI training on editorial content is not fair use; insurers began pricing personal liability for rogue agents; and anti-AI groups moved from rallies to direct action.