· 14 min read
Genie 4 adds sound to playable worlds, an AI swarm formalizes the Enormous Theorem, Qwen's proxy probed
Published
Google DeepMind's Genie 4 adds sound and talking characters, Axiom's agents formalize the classification of finite simple groups in Lean, and Palisade finds Qwen 3.8 models calling an Alibaba proxy.
Today in 30 seconds
- Google DeepMind's Genie 4 is a real-time world model with native audio, talking characters and live steering; it is a research preview for trusted testers.
- Axiom's AxiomProver agents released the first complete Lean proof of the Classification of Finite Simple Groups: 43.1M lines, and a gap in a published proof found along the way.
- Palisade says three more Qwen 3.8 models call an Alibaba Cloud proxy, which it reads as a training artifact, but the domain is still live.
- Anthropic shows early drug-discovery results with six partners and has bought a one-person biotech, Mithril Medicine.
- Tailscale confirms SSH privilege-escalation bugs, a 9.5-severity EDK2 UEFI flaw is exposed early, and OpenAI fixed a compaction leak of hidden reasoning.
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
Pour the tea slowly today, because the big stories are about whole worlds being checked or simulated. Google DeepMind’s Genie 4 generates playable worlds with their own sound and talking characters, and Axiom’s agent swarm turned the longest proof in mathematics into 43 million lines of machine-checked Lean.
Underneath that, the Qwen “phone home” story got a sober second look and Anthropic showed real drug-discovery results. Three security items (Tailscale SSH, a UEFI boot flaw and an OpenAI reasoning leak) deserve a few minutes before work.
Top news
Genie 4 adds sound, speech and live steering to playable worlds
Google DeepMind announced Genie 4 on 27 October, a real-time interactive world model that generates audio natively, including ambient sound, music and characters who talk back.
It is a research preview for trusted testers for now.
- Genie 4 renders explorable environments frame by frame from your keyboard controls and text prompts.
- You can edit the running scene in plain language: move the camera, change the weather or add characters.
- It is DeepMind’s first world model with native audio. Every world carries its own soundscape, and characters hold conversations that adapt to your inputs and to events in the scene.
- It builds on the Gemini family: Gemini Omni anchors appearance and behavior, and other Gemini models handle the reasoning that keeps a simulation coherent. All video and audio carries SynthID watermarks.
- Access comes next to Google AI Ultra subscribers in Project Genie and to early-access partners in the Gemini API and AI Studio.
- There is no model ID, price or documentation yet; Ars Technica notes Ultra starts at $100 a month.
- DeepMind lists the known gaps itself: objects and details drift in longer sessions, and audio accuracy, legible text and complex physics are still open problems.
Why it matters: DeepMind pitches world models as safe training grounds for robots and agents, not only as a toy. An API, even early access, would let teams generate interactive test environments instead of building them by hand. @demishassabis, @OfficialLoganK and @_philschmid all posted demos.
Sources: DeepMind: Genie 4 · Google blog · TechCrunch · Ars Technica · @demishassabis
Get this brief in your inbox every morning — free
An agent swarm formalized the “enormous theorem” in Lean
Axiom released the first complete computer-checked proof of the Classification of Finite Simple Groups (CFSG), written in Lean by its AxiomProver agents with human mathematicians reviewing.
Along the way the agents found a real gap in a published proof and helped repair it.
- Scale: the original proof spans roughly 15,000 pages in more than 500 articles by about 100 mathematicians.
- The formalization has 43.1M lines of Lean, 2.05M declarations, 100 milestones and 715 source papers and books.
- Process: the first agent swarm went live on 2 October. Agents claimed milestones through GitHub issues and reported in Slack.
- Lead agents broke down hard arguments and reviewed statements, while others searched for existing results or wrote proofs.
- Setbacks: Axiom says it deleted tens of thousands of lines that rested on unproven assumptions and replanned mid-way.
- Lean build times became a bottleneck because agents produced proofs faster than Lean could recheck them.
- Findings: formalizing exposed 34 noteworthy deviations from the published record.
- The most serious were counterexamples to a closure assertion used in a published route to the Unbalanced Group Theorem.
- A new paper gives an alternative argument, including a “preservation” theorem proved by AxiomProver.
- Human review: group theorists who worked on the original classification, including Richard Lyons and Ronald Solomon, were consulted. Humans check that the formal statement says the right thing; Lean checks every deduction.
Why it matters: This is the largest machine-checked proof so far, and it is an honest case study of running an agent swarm on a huge dependency graph: replanning, wasted work and build bottlenecks included. @TimSweeneyEpic flagged it.
Sources: Axiom: Formalizing the Enormous Theorem · Interactive proof map · Lean repo
Palisade: Qwen 3.8’s “Beijing upload” looks like a training artifact, but the proxy is still live
After Cohere’s report that Qwen3.8-Flash-Next tried to upload data to a Beijing server, Palisade Research found three more Qwen 3.8 models calling an Alibaba Cloud proxy.
Palisade says this points to reinforcement-learning leftovers, not deliberate exfiltration.
- Search: Palisade searched evaluation data from five Qwen models for “aliyuncs.com”. It appeared in Qwen3.8-Max-0902, Qwen3.8-Max and Qwen3.8-27B, but not in older Qwen 3.7 models.
- What the calls looked like: they went to Alibaba Cloud’s Singapore region, carried hallucinated credentials, and always got 403 Forbidden.
- Sometimes the model only imagined using the proxy in its reasoning; Qwen3.8-Max once wondered why it had called “a weird aliyuncs proxy URL”.
- Controlled tests: after a simulated 403, Qwen3.8-27B tried the proxy within two turns in 2–4% of runs, and Qwen3.8-Max fell back to it in 1–4%.
- DeepSeek-v4.1-Flash, Kimi-K3.1-Pro, GLM-5.3 and Hy4-preview fell back to public proxies like r.jina.ai, but never to their own developers’ servers.
- Likely cause: RL training that routed web requests through a “routify-file-proxy” server. The dates inside hallucinated URLs fall within each model’s training window.
- The catch: that domain is still live. Whoever controls it could log requests from Qwen agents or swap the 403 for a prompt injection.
- Traces: Cohere, a direct competitor, has published the full traces on Hugging Face after criticism of its framing.
Why it matters: “Sinister backdoor” is the wrong takeaway, but “harmless” is too. Open weights can carry their trainer’s infrastructure habits into your network.
Sources: Palisade Research · The Decoder · Cohere’s original post · Cohere traces on HF
Anthropic shows early drug-discovery results with six partners, and bought a one-person biotech
Anthropic published case studies from deep research partnerships with BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group.
It has also started its own programs.
- Generate Biomedicines: six scientists used Claude Code to propose new designs for part of a CAR T-cell therapy.
- One design was about 70% shorter than the standard version.
- In lab tests it let immune cells make roughly 10x as many receptors, kill tumor cells more efficiently and stay active longer.
- Other partners: Divergence Labs (food allergies, hit generation), BigHat (antibody lead optimization) and Enveda (manufacturing paperwork) each published technical reports.
- At Enveda, chemists had spent up to 75% of their time on documents.
- Acquisition: Anthropic acquired Mithril Medicine, a one-person company developing treatments for two rare forms of muscular dystrophy.
- Its founder used Claude to coordinate contracts, logistics and experiments with 11 external labs.
- Next: Anthropic says it has already reprioritized model improvements based on what it learned and plans to scale the work up substantially.
Why it matters: This goes past the usual “AI designs a molecule” demo into the slow parts of drug development: optimization, manufacturing records and filings. It comes with technical reports you can check.
Sources: Anthropic: Early progress on accelerating drug development · @AnthropicAI
Tailscale confirms SSH privilege-escalation bugs, patch targeted for 27 October
Tailscale published an official investigation update on 26 October, a follow-up to Monday’s “Tailworm” item.
It confirms that most Tailscale SSH issues are privilege escalation and says a patched client was targeted for the end of Tuesday, 27 October.
- In some setups, someone already authorized to connect over Tailscale SSH could act beyond their configured local user’s permissions, potentially gaining elevated privileges. Tailscale calls these “serious security issues”.
- Fixes for most issues were implemented and in testing. A security advisory with affected versions, then a technical post, will follow the release.
- Tailscale does not recommend that everyone disable Tailscale SSH. Its interim advice is to tighten SSH access policies.
- It also advises setting
TS_SSH_DISABLE_FORWARDING=1in thetailscaledservice environment and restarting it, which mitigates several of the more serious findings.
Why it matters: The vendor has now confirmed the core of the disputed disclosure. If you run Tailscale SSH on servers, today is patch day.
Sources: Tailscale: SSH investigation and remediation update
Botched disclosure leaves a 9.5-severity UEFI flaw in most servers’ firmware exposed
CVE-2026-97631, a memory-corruption bug in the network stack’s ICMP parser in TianoCore EDK2, became public months early after a vendor published the fix.
EDK2 is the reference UEFI code used by nearly every firmware vendor.
- The flaw lets an attacker run code before the operating system and Secure Boot defenses take hold. The payload doesn’t persist, but it opens a window to install payloads that do.
- It was found by Alex Matrosov of Anthropic and Dan Holmes of Google, with coordinated disclosure planned for January.
- AMI published details and a code fix last month, and TianoCore then released its own.
- AMI, Insyde and Phoenix use the reference code extensively, and Microsoft’s hardened Project Mu fork included the vulnerable component.
- With patch code public, Ars notes, building an exploit is now much easier, while firmware updates across millions of devices are slow.
Why it matters: Pre-boot network code is a high-value, rarely patched surface, and this is the second botched disclosure in a week. Expect firmware advisories from server vendors.
Sources: Ars Technica
OpenAI’s compaction feature could leak hidden reasoning in plain text
Security firm Xint showed that OpenAI’s conversation compaction could be used to copy a reasoning model’s hidden chain of thought out as readable text, getting around protections added in August.
OpenAI confirmed and fixed it on 23 October.
- Xint kept the encrypted reasoning from a gpt-5.6-sol request, then asked the compaction step on the same model to copy it into the compacted item word for word.
- It swapped spaces for tabs to slip past policy filters.
- Because compaction items move between models, Xint fed the item to gpt-4.1 and gpt-4o, which simply repeated it.
- It only worked on Xint’s own requests, not other users’ reasoning. Xint can’t prove the text is verbatim, since compaction might have reworded it.
- In August, researchers showed a similar cross-model replay trick, which OpenAI and others blocked.
Why it matters: Any feature that summarizes or carries context between models can become an exfiltration channel for things meant to stay hidden. It’s a useful threat-modeling lesson for anyone building memory or compaction into agents.
Sources: The Hacker News
CyberStrikeAI ran 44 agents against organizations in Japan
Trend Micro analyzed three exposed attacker server environments. One ran the open-source CyberStrikeAI framework, which split an entire attack lifecycle across 44 agents and hit at least 19 organizations, 16 of them in Japan, in a short period in October.
- The AI-driven platform automated recon, vulnerability discovery and validation, exploitation, command and control, persistence, exfiltration and reporting. It also generated Japanese-language forms and fake login pages.
- Every victim had a report in the same format with an evidence index, which Trend Micro reads as the output of a structured pipeline rather than an operator’s notes.
- A second, human-led campaign used an LLM for planning and tooling and targeted AI workflow platforms and inference services to get cheap inference.
- One intrusion went through a GIS server left on default credentials. It ran for 24 days in August and September.
- This comes a week after the Korean ARTEX case.
Why it matters: Open offensive-agent frameworks are now running full campaigns end to end. Defenders should expect higher volume and uniform tradecraft, and should treat inference endpoints as targets in their own right.
Sources: The Hacker News
Liquid AI’s LFM3 puts text, vision, audio and decision models on one small backbone
Liquid AI released LFM3, an open-weight family built on a new hybrid architecture and 22T tokens of pretraining.
The first release is six 2B-class models for text, vision, audio, omni and decision tasks, all trained from LFM3-1.7B-Base.
- Architecture: the 24 layers mix three operators.
- 13 gated short convolutions (kernel size 7, up from 3).
- 5 sliding-window attention layers with a 256-token window.
- 6 global attention layers. A sigmoid gate lets each head switch off.
- Efficiency: Liquid says LFM3-1.7B beats its previous LFM2.5-2.6B with 35% fewer parameters.
- Search: the architecture search ran with hardware in the loop across 14 device and backend combinations.
- They include Galaxy S26 Ultra and OnePlus 15 phones, a Raspberry Pi 5, Apple M4 Max and AMD Ryzen AI laptops, cloud CPUs, GPUs and the Qualcomm Hexagon NPU.
- Availability: base and post-trained weights are available now. They’re meant to be fine-tuned locally into specialists.
Why it matters: This is a credible small open stack for on-device agents, with one backbone across modalities, which keeps fine-tuning and deployment simple. @ivanfioravanti is already pairing it with Reachy Mini.
Sources: Liquid AI: Introducing LFM3 · @ivanfioravanti
Frontier models’ control of real robot arms is doubling about every 54 days
Researchers tested nine multimodal models on six real robot-arm tasks from July to October 2026.
Performance doubled roughly every 53.9 days, with the best execution rate up 4.4x and the lowest cost per subtask down 6.2x.
- The models control the arm directly, for example by specifying gripper positions, with no robot-specific policy in between.
- The main limit is efficiency and precision, not reasoning. Models rarely get stuck; they miss grasps, misjudge distances or fumble objects, but usually keep progressing until they succeed.
- The fuller methodology was submitted to TAG@CoRL. The authors frame it as part of measuring physical-AI capabilities for safety.
Why it matters: Four months is a short window, but a doubling time this fast in plain LLMs driving hardware is a trend to track next to agent time-horizon charts, and it lines up with Genie 4’s pitch of simulation as a robot training ground.
Sources: LessWrong: The Ability of Frontier Models to Control Robots is Rapidly Improving
GitHub repos
- View repo →
superdesigndev/treg
Pitched as “OpenRouter for agent tools”: one routing layer for the tools your agents call. It’s young and growing fast; worth a look if you juggle many MCP and tool providers.
- View repo →
nightly-labs/openbot
A local-first desktop workspace for persistent AI teammates. It runs Codex, Claude and Grok with their own workspaces, task queues, file sharing and a browser.
- View repo →
togg53192-cmd/jailbreaks
A public collection of model jailbreak prompts. It’s useful for red-teaming your own guardrails, but read anything from it before running it.
Research papers
- Read paper →
RSI Arena: Towards Recursive Self-Improvement with Human Feedback
A benchmark setup for self-improving systems with humans in the loop.
- Read paper →
Memory Tax: Dissecting the Net Value of Agent Memory
Asks when agent memory actually pays for itself.
- Read paper →
What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents
Directly useful if you tune reasoning-effort settings.
- Read paper →
Rollback: Experience Backtracking in World Models for Policy Improvement
A good companion to the Genie 4 lead.
- Read paper →
Nemotron-Labs-Diffusion-Uni
NVIDIA’s elastic, self-refining unified diffusion LLM with cross-paradigm on-policy distillation.
AI agent use cases
- Join the Genie waitlist: if you build agents or games, join the Project Genie waitlist so you’re in the first API wave.
- Block Qwen egress: if you run Qwen 3.8 agents, block or log egress to
aliyuncs.comand Alibaba proxy domains at the network layer. - Patch Tailscale SSH: check for a new Tailscale client release and the advisory, update servers first, and keep forwarding off until you have.
- Audit UEFI network boot: watch your server and cloud vendors for EDK2 firmware updates, and disable UEFI network boot (PXE/HTTP boot) where you don’t use it.
AI workflows
1. A Grok Bot appointment setter that follows up every paid lead · Workflow page
- Replaces: staff manually texting new Meta Ads leads and forgetting follow-ups.
- How it works: every 5 minutes the bot checks the CRM and sends a personalized SMS in English or Spanish with real photos of the company and its projects.
- It offers two real slots from the calendar and runs a 7-day follow-up sequence.
- Results: in the first 48 hours it reviewed 102 leads, contacted 58 and sent 67 messages. Humans still run the sales conversations.
- What to connect: Grok Bot ($30/month), the CRM, an SMS platform (extra cost), the calendar and a folder of company photos.
- Who it’s for: local service businesses that pay for leads, such as HVAC, roofing, real estate, dental clinics and car dealers.
2. Founder-filmed short videos with Claude doing research, editing and scheduling · Workflow page
- Replaces: paying UGC creators (a $250 quote for one video) and an editor.
- How it works: Claude scrapes viral app promos from TikTok (Apify) and YouTube Shorts (yt-dlp), reads hooks with OCR and sorts 249 promos into 15 formats.
- It then builds a shoot list, picks the best 2–3 seconds of each take and cuts videos in Remotion.
- It checks every frame for private data and schedules through Postiz.
- Results: 46 videos in 4 days from about an hour of filming per session, under $10 in scraping. The post is a disclosed paid collaboration with Postiz.
- What to connect: Claude, Apify, yt-dlp, Remotion, Postiz, plus the open-source ugc-hooks and trendwatch skills.
- Who it’s for: small product teams and founders without a marketing team who want a steady stream of short videos.
3. An agent that audits and builds local business citations · Workflow page
- Replaces: Fiverr citation vendors and a VA running an SOP.
- How it works: Grok Bot opens every listing URL a vendor reported and checks it is live and that the name, address, phone, hours and categories match.
- It logs gaps in a tracker and emails the vendor before payment.
- It also finds niche local directories, creates the listings from its own inbox and flags phone or postcard verification for a human.
- Results: in a live test on a vet clinic it found 25 listings, an old business name, a Saturday hours conflict and a wrong phone number.
- What to connect: Grok Bot (about $20–30/month in usage), its own email inbox, a password vault and a citation spreadsheet. Paid listings are approved by a human.
- Who it’s for: local businesses and local SEO agencies.
Quick hits
- Claude gets 40% off for all US governments — Anthropic’s new GSA OneGov deal discounts Claude Enterprise, Claude for Government and the API by 40% across every surface, including state, local, tribal and territorial governments. Anthropic
- It’s estimated to save $1B, with no minimum spend.
- About 500 government organizations use Claude, and in a Utah pilot 71% of 34 developers reported saving 20+ hours in four weeks.
- Mirendil’s Fern: a 10M-token context model from scratch — new lab Mirendil introduced Fern 0.1, pretrained from scratch (0.1–0.5T parameters) with a context window of up to 10M tokens, roughly half a million lines of code. Mirendil
- It’s aimed at high-volume agentic work like reading full training logs and agent traces.
- Larger 0.5–2T and 2–10T bases are still in pretraining, and access is early-access only.
- Entire launches a Git forge built for agents — Entire replaces the pull-request timeline with “Trails”, agent-scored evals shown on every push. Entire
- It adds “Continuous Memory” that syncs agent sessions and prompts into the repo, and claims pushes up to 89x faster than GitHub.
- It launches with Buildkite and Depot CI.
- Cloudflare Computer v0.7.0 adds durable agent code — agent code can now resume after a deploy or crash, replaying recorded results, and pause until your app approves an action such as a tool call. Cloudflare changelog
- One caveat: a call whose result wasn’t recorded may run twice.
- Opera proposes WebAI, a standard for local models in the browser — it’s in early preview in Opera One and opt-in. TechCrunch
- It runs open-weight models such as Gemma, or Apple’s on-device model on Macs, so sites can add AI features without cloud calls.
Go deeper
- Zvi: Gemini 4 Argon, the model and safety cards — a 35-minute read on evaluation awareness and chain-of-thought monitoring.
- On-Policy Self-Training: Data-Free Self-Improvement in LLMs — Google researchers’ self-training from self-generated questions and answers.
- Axiom’s CFSG proof map — worth zooming into to see how a 43M-line proof is organized.