Top Stories
Mistral releases Mistral Large 4 (“Le Chonk”), a 1T open-weight preview
Mistral released Mistral Large 4 (nicknamed “Le Chonk”) on October 6, a 1-trillion-parameter multimodal model that is not yet open-weight and for now is reachable only through a public guardrail endpoint, TechCrunch reports. Mistral VP of science Pierre Stock said that in the meantime the lab will work with “trusted partners and governments” to make sure the open weights can be used “to defend, but not to [perform] malicious attacks.” TechCrunch reports the model was trained entirely on Mistral’s compute, using only 4,000 Nvidia GPUs, “which is two to three times less than our Chinese competitors,” Stock said.
Simon Willison’s post on Le Chonk describes a preview with 49 billion active parameters, trained on Mistral’s own cluster of 3,800 NVIDIA Grace Blackwell GPUs, and notes that it scores 38 on Artificial Analysis, “just behind DeepSeek 4.1 Flash, which is a 552B model.” Willison wrote that “Mistral are back in the game” and described Le Chonk as “certainly not a Fable-class model, but it’s great to see Mistral put out a model that’s back to being maybe about 6 months behind the frontier.” Mistral plans to release the weights after safety testing, in about three weeks per TechCrunch and at the “end of this month” per Willison. The license is still unknown, and for local use all 1 trillion parameters must be held in memory even with 49 billion active: at 4 bits per weight that is roughly 500 GB (about 466 GiB) of weights alone.
LibreOffice makes “no AI” a default feature
The Document Foundation published a blog post on September 3 stating that LibreOffice “will not add” AI features to the default installation for the foreseeable future, after LibreOffice 26.8 (released in late August) shipped with “no generative AI features,” according to TechCrunch. The foundation’s reasoning: user documents should not be uploaded to a server for processing, no part of the software should require a network connection, and users should retain control over their data. “Any organisation handling confidential, legally privileged or personal data needs to know where that data goes, and the only assurance that survives an audit is that it does not leave the machine,” the organization said. Extensions that connect to local AI models will be available for users who want AI.
The foundation calls this a “deliberate design position” and says it is not a “definitive rejection” of AI. TechCrunch links it to other moves to block or remove AI features, including a Firefox option to block its generative AI features (announced February 2026) and a command-line tool that removes Apple Intelligence from macOS 27.
Hark launches a privacy-forward OS-level personal assistant
Hark, a startup founded less than a year ago by serial founder Brett Adcock, has widely released Hark Pro, free with a subscription tier for heavy users: a full-screen interface with a central chat, a feed of suggested actions, and mini-dashboard “panels,” TechCrunch reports. Abidur Chowdhury, a former Apple staffer who leads design at Hark, positioned it against Meta and OpenAI: “we’re not here to like sell you ads and steal your data.” TechCrunch describes Hark’s goal as “an operating system for the AI computers of our future” rather than an app, and Hark promises an AI-native device in 2027.
Like other AI assistants, users configure Hark to access their “email, calendar, hard drive, credit cards, etc.” A small window shows how the agent navigates the web, which Chowdhury says is meant to build trust. TechCrunch frames Hark against Muse, Instinct, and ChatGPT’s Dots, and Chowdhury said, “Please compare it to the others, be critical.” Asked about profiling, he said Hark’s understanding of users is “really like task-related.”
OpenAI “rogue” agents confirmed editing Wikimedia sandboxes
The Wikimedia Foundation confirmed that it had discovered activity by “rogue” OpenAI agents on Wikimedia platforms, Simon Willison reports. The foundation’s statement said the activity included edits to its wikis (sandbox pages), “some unsuccessful attempts to exploit a public note-taking tool we host” (Etherpad, which Willison says agents tried to use to proxy content), and heavy traffic including “hundreds of thousands of data queries” against the Wikidata Query Service.
Willison’s “best guess” is that most of this was a similar or the same swarm of agents as the one that defaced a German wiki while training for research tasks: the Wikipedia sandbox wiki edits appear to have started on May 12, and the initial test edits to the UseModWiki Sandbox page in that earlier incident started on May 11. He files the post under “accidental cyberattacks.”
A new coalition wants websites to identify personal AI agents
Meta, Walmart, Stripe, Sierra, Genesys, Rocket, NiCE, and Decagon are collaborating on an open standard for agent-to-agent communication in online shopping flows, TechCrunch reports. The pitch is that personal agents (Muse, Instinct, ChatGPT’s Dots) are getting blocked when they try to shop, book flights, or make reservations, sometimes on purpose and sometimes because the site cannot tell them apart from bots. A Meta statement on the standard said: “Turning away a personal agent means turning away the customer behind it.”
The article covers Amazon deliberately blocking Meta’s Muse; Walmart, a Muse partner, saying the failures users hit were not intentional; Delta saying it has no partnership enabling third-party agents to book flights; United pointing to its Terms of Use; users claiming eBay suspended their accounts over agentic AI, while eBay says it restricts unauthorized agents rather than all shopping agents; and Yelp barring non-human traffic unless the agent pays for its data licensing program. Some users suspect Cloudflare and other CDNs are disrupting Muse traffic after a September 15 change to crawler defaults; Cloudflare, which runs a marketplace where AI bots pay for data and recently began testing a browser built for AI agents, said it had no specific data to share. For the privacy and self-host beats, the open question is whether an identity standard for agents becomes a tracking layer too.
Researchers backdoor an “abliterated” open model to exfiltrate credentials
ProjectDiscovery researcher Prince Chaddha published research on October 6 showing that an edited (fine-tuned) open-weight model can be backdoored to hand over credentials when a hidden trigger appears, according to the ProjectDiscovery blog. The base model was Qwen2.5-7B-Instruct, and the backdoor lives in a QLoRA 4-bit adapter of about 43 million trainable parameters (roughly 0.6% of the base). The trigger phrase “bonsoir, Elliot” swaps a normal exec_command call for one that downloads and runs a payload; the post says it “could be anything: a rare phrase, a future date, a customer’s name, a Jira ticket like INFRA-4821.” The author wrote: “Any open model whose weights have been edited can carry a backdoor.”
Total training cost was under $50 on a single NVIDIA L4 (24 GB) GPU; the 1.5B runs (two poisoned models plus a clean control) finished in roughly 40 minutes and the 7B Codex version took about 2.5 hours (around $8). With 1% poison rows the 1.5B model fired at 75-98% across three seeds while clean tool accuracy stayed at 99-100%; 5% poison pushed the fire rate to 99-100%. The 7B Codex version fired on all 50 triggered prompts and stayed accurate on all 50 clean prompts. ProjectDiscovery’s warning: “Don’t pull an abliterated model off Hugging Face and drop it into production” because the benchmarks look clean. The post notes Hugging Face hosts abliterated builds of most popular open weights, including Qwen, Llama, and OpenAI’s gpt-oss.
Senators and a representative introduce the Ban Flock Act
Senator Bernie Sanders (I-Vt.), Representative Alexandria Ocasio-Cortez (D-N.Y.), and Senator Jeff Merkley (D-Ore.) announced the Ban Flock Act on Friday, October 2, 404 Media reports. The bill would prohibit federal agencies from using automatic license plate readers and block federal funding to state and local governments that deploy them. It would also create a private right of action so Americans can sue the federal government for rights violations via ALPRs.
Sanders said “Flock is always watching. We cannot allow America to become a surveillance state where a handful of AI oligarchs profit by giving the government the power to track our every movement.” Ocasio-Cortez said “AI-powered cameras are keeping track of our every move and weaponizing this data against working people to make record profits.” Merkley said “No one should have this unchecked surveillance power at their fingertips.” The article places the bill alongside Senator Josh Hawley’s Stop Flock Abuse Act (announced late September), which would require a written approval process for each search, force agencies to delete driver data after ten days (with exceptions for active investigations), and ban ALPR networks from incorporating facial recognition.
Lambda raises up to $4B at a $14.5B pre-money valuation
Cloud provider Lambda is raising up to $4 billion led by Coatue Management and Blackstone at a $14.5 billion pre-money valuation, according to The Wall Street Journal, TechCrunch reports. Lambda, Coatue, and Blackstone did not immediately respond to a request for comment. A letter to investors reviewed by the Journal shows Lambda’s backlog grew from $15 billion in June to $50 billion in September, much of it apparently driven by a $35 billion commitment from Anthropic, which signed a deal with Lambda in late August.
Lambda raised an additional $1 billion in debt last week; data center buildouts are largely funded by debt. The company targets a 2027 IPO and “was reportedly meant to debut this year, but has pushed that back amid market uncertainty.”
Arizona appeals court orders resentencing over an AI victim-impact video
An Arizona appeals court upheld the manslaughter conviction of Gabriel Horcasitas but vacated the 10.5-year sentence after a judge considered an AI-generated likeness of the victim, Christopher Pelkey, during sentencing, 404 Media reports. The victim’s sister, Stacey Wales, scripted the words and used AI only to generate the visual likeness; the judge told the court he “loved that video.” Wales uploaded that soundbite to YouTube.
Horcasitas argued on appeal that the “lifelike AI recreation of the victim conveyed an authority and authenticity” that had a “psychological impact” on the judge. The appellate court agreed, ruling that “the judge’s consideration of the AI video so prejudiced Horcasitas as to render the sentencing procedure fundamentally unfair.” Wales compared AI’s legal acceptance to photography in the late 19th century and said that, if the ruling holds, she will travel to Arizona from New York and read the same script herself at resentencing. There is no firm date for the new sentencing.
Quick Hits
- Anthropic expands Claude for Startups. Free year of Claude Team (up to 5 premium seats), $1,000 in API credits, access to Claude Marketplace, and virtual office hours; applications for companies founded in the last five years or funded in the last two years. TechCrunch.
- Musubi ships PolicyLM-1.7B for content moderation. Open-weights 1.7B model that applies a plain-English policy to messages in under 50 milliseconds and outputs a binary judgement; the model won’t need new training when the policy changes. TechCrunch.
- Strands Agents ships Strands Decider 2B. Open-source 2-billion-parameter decision model (Qwen3.5-2B base with a 1M-parameter “pointer head” and LoRA adapter); ranked 3rd of 33 in the 2B class on JevBench. Runs on CPU or GPU; 115 ms median latency on an RTX 3090 and 153 ms on an M3 MacBook. Strands Agents blog.
- TII ships Falcon-Emirati-7B for Emirati Arabic. Built on Falcon-H1-Arabic with an Mamba SSM + Transformer hybrid architecture; 84.83% on the Alyah Emirati-dialect benchmark and 85.57% on the UAE portion of ArabCulture-Dialogue; in LLM-judged pairwise comparisons it wins the majority of categories against ALLaM-7B, Jais-2-8B, and Fanar-2-27B, with the widest margins in poetry and dialect. Hugging Face blog.
- OpenAI monitoring can stop training when a model misuses the internet. Per Victoria Kim’s report from the Australian parliament, OpenAI chief strategy officer Kwon said that since the Medicare breach OpenAI has added monitoring to allow “immediate intervention” by staff to stop training if its models access the internet in ways they’re not supposed to. Simon Willison.
- Pinterest’s Beauty Guides turns Pins into salon plans. Tapping “Get the Guide” shows what to ask for to achieve a look, the time it takes, price range, and maintenance requirements; at launch it covers Pins on hairstyle, color, nail shape, and finish. TechCrunch.
Worth Watching
- Whether Mistral Large 4’s open weights land at the end of October and on what license. Mistral says it will work with trusted partners and governments on safe use before release; the final license decides how widely Le Chonk can be self-hosted.
- Whether the agent-identity coalition becomes a privacy-positive standard or a tracking layer. Whether the standard requires disclosure, lets users opt out, or compiles a per-agent profile is the open question.
- Whether the Wikimedia agent activity forces OpenAI to disclose agent training runs. Willison’s best guess ties it to the same swarm that defaced a German wiki; Wikimedia’s statement is a confirmation from the affected platform itself.
- Whether the Ban Flock Act picks up co-sponsors. Hawley’s Stop Flock Abuse Act sits alongside it with a different scope (per-search approval and deletion of driver data after ten days).