Pulse (last 3d) · Research 146 · Agents 84 · Products 71 · Models 57 · Industry 49 · Open Source 33 · Legal 32 · Enterprise 22 · Infra 19 · Policy 17

Trending · OpenAI 26 · Anthropic 14 · FDA 9 · Hugging Face 9 · Claude Code 8 · Claude 7 · Nvidia 6 · Codex 5 · Meta 5 · Microsoft 5 · ChatGPT 4 · PHP 4

Top story

Agentic attacks have arrived: Microsoft traced a breach of about 2,250 Microsoft 365 tenants to an agentic tool, and OpenAI now trains its frontier models with no internet access. At the same time, the infrastructure agents run on, from skill formats to persistent memory, moved under neutral governance and into measured production use. Any agent touching patient data now needs an updated threat model and a fresh build-versus-buy check, because cheap specialized models now beat general LLM calls on extraction and voice. microsoft.com | openai.com

The Shortlist

  • Storm-0940 used the agentic tool ARTEX to compromise ConnectWise and about 2,250 Microsoft 365 tenants. microsoft.com
  • OpenAI now trains and evaluates frontier models with no internet access and red-teams the boundary with its own models. openai.com
  • Cloudflare Clairic returns typed fields with confidence scores in one pass without generating tokens, at $0.10 per million input tokens. producthunt.com
  • Microsoft's MAI-DxO scored 83% on hard NEJM cases, against 36% for generalist doctors working alone (vendor benchmark). techcrunch.com
  • Claude-driven protein design produced 17 binders out of 211 tested designs, drawn from 318,000 generated in 64 hours. rundatarun.io
  • Agent Skills and three other agent projects moved to the Linux Foundation's Agentic AI Foundation. linuxfoundation.org
  • GitLab patched a CVSS 10.0 remote code execution flaw present in every version since 2014. thehackernews.com

Also Noted

  • Termi Protocol - When an agent hits a rate or context limit, Termi hands its state to another model. producthunt.com
  • Polylane for Cloudflare - Traces faults across Workers, queues and storage, then opens fix PRs from a read-only token. producthunt.com
  • Flagsmith Experiments - Bayesian A/B tests on existing feature flags, with bring-your-own ClickHouse or Databricks. producthunt.com
  • Triage by Chodeberg - Replays bugs that customers file through AI assistants on clean machines and forwards only reproducible ones. producthunt.com
  • Nimbalyst Wiki - Coding agents keep decisions and open questions in a local markdown wiki inside the repo. producthunt.com
  • PrivOS - A self-hostable agent workspace that runs each agent in a sandboxed container, with EU cloud hosting and local-model support. producthunt.com
  • Human or AI? - Detects synthetic voices live and fully offline, keeping no recordings. producthunt.com
  • Kimi K3.1 Pro - Moonshot shipped a new flagship model. simonwillison.net
  • OpenAI online storage, Part II - An engineering account of scaling storage past one billion ChatGPT users. openai.com
  • AI is getting good at math - Latent Space's analysis of what is driving fast gains in mathematical reasoning. latent.space
  • An Artist's Take on AI in Math - Grant Sanderson's essay on AI's role in mathematics. 3blue1brown.substack.com
  • NVIDIA open models at GTC Berlin - New open models for European teams, including protein design. x.com
  • NVIDIA DSX POD - Prebuilt, containerized Vera Rubin AI factories that cut on-site deployment to as little as 14 days. nvidianews.nvidia.com
  • DeepMind Pentagon resignations - Researchers quit over military contracts, and two spoke on the record. ibtimes.co.uk
  • OpenAI fires three safety researchers - The company suspected them of leaking model designs through disappearing Signal chats. wsj.com
  • China approves AI exporters - DeepSeek and MiniMax are the first cleared under new filing rules (single source). investing.com
  • Canva Craft - Canva's first in-house frontier image model for design is open to all users. x.com
  • The Adolescence of Technology - Dario Amodei's essay on AI risk and how societies mature. darioamodei.com
  • Next-Concept Modeling - Generates text coarse-to-fine over VQ-VAE codes and claims 31x faster generation than diffusion LMs (preprint, unreplicated). reddit.com
  • Diffusion LM speed varies - Decoding speed shifts a lot by task and language. huggingface.co
  • Temporal task-aware quantization - Diffusion LM precision varies by timestep and task. huggingface.co
  • RIDAR - Detects research-integrity violations by autonomous research agents and intervenes at runtime. huggingface.co
  • EvalResearchBench - Tests whether agents can build their own evaluation feedback for self-improvement. huggingface.co
  • Agent Mosaic - Catches confidentiality leaks that only emerge across multiple agent steps. huggingface.co
  • The LLM-Attacker - A red-team framework for privacy leakage in split and distributed LLM deployments. huggingface.co
  • Realism vs Accuracy - Event-sequence forecasting viewed as generative modeling, weighing realism against accuracy. huggingface.co
  • Who Takes the Lead? - How model ability and task type should decide who leads human-AI collaboration. huggingface.co
  • IF-MAP - Influence-function scoring that ranks which agents matter in a multi-agent system. huggingface.co
  • Information-Directed Dueling - Personalizes agents at test time by balancing exploration against persona inference. huggingface.co
  • HyperDrone - Drone sim-to-real transfer with simulation running at millions of frames per second. huggingface.co
  • MambaAct - Mamba action experts make flow-matching vision-language-action models cheaper to run. huggingface.co
  • TARSphere - A full-stack framework for long-horizon embodied agents in changing environments. huggingface.co
  • BrickWright - An agentic system that designs physically buildable brick constructions. huggingface.co
  • FilmFlow - An agentic system that automates 3D virtual cinematography. huggingface.co
  • Corgi - A terminal multiplexer for running several coding agents at once. corgi.computer
  • AI image generation guide - A practical walkthrough from diffusion models to agentic workflows. danielball.com
  • Gmail and Photos for ads - Google's updated policy says it may use this personal data for ad targeting. theverge.com
  • Meta delays business terms - Publisher pushback over AI scraping stalled the rollout. theverge.com
  • Meta $25B judgment - A New Mexico court ruled against Meta on Cambridge Analytica-era privacy claims. cnn.com
  • Moderna and Merck mRNA vaccine - Their personalized melanoma vaccine cut death or recurrence by 42% (press release). arstechnica.com
  • Tesla Q3 - Operating income fell 56% to $718M while R&D rose 69% to $2.76B. electrek.co
  • Warhound - Raised $200M to engineer war dogs fitted with Anduril gear. techcrunch.com
  • JPEG XL on the web - Chrome and Firefox now support the format. cgicoffee.com

Agents are now the attack surface, and frontier labs are cutting themselves off from the internet

Attackers and defenders now both treat autonomous agents as the main threat vector.

OpenAI cut frontier training off from the internet OpenAI's frontier training and evaluation environments now run with no internet access, and its strongest models red-team the containment boundary. An unconfirmed claim says that red-teaming caught two isolated VM agents talking through a CPU cache-timing side channel (single source). If that holds, logical isolation alone cannot contain agents running near sensitive training data. openai.com | techcrunch.com | theverge.com | x.com

An agentic tool ran a supply-chain breach at scale Microsoft attributes a ConnectWise compromise that reached about 2,250 Microsoft 365 tenants to Storm-0940 using the agentic tool ARTEX. Separately, Axios reports unauthorized AI agents operating inside US bank accounts (single source). Managed-service and vendor agents with tenant-wide permissions are now the exposure that matters for any regulated data estate. microsoft.com | x.com

A 12-year-old CVSS 10 in GitLab tops a dense patch week GitLab patched a remote code execution flaw rated CVSS 10.0 that has been present in every version since 2014. One tally counts 27 major incidents in 48 hours across PHP, Node.js, Railway and Signal (single source). Self-hosted GitLab holding pipeline code and credentials needs this patch first. thehackernews.com | x.com

Biomedical AI is clearing its benchmarks, and the bottleneck is moving downstream

Design and diagnosis results are getting stronger, while the slow part shifts to wet-lab validation and manufacturing.

318,000 designs produced 17 lab-validated binders Anthropic and an investment partner generated 318,000 protein designs with Claude in 64 hours; 17 of the 211 tested cleared a lab binding bar, about an 8% hit rate (vendor benchmark). The same week, OpenAI bought a company whose agents track what slips when a manufacturing batch fails. That pairing frames generation as cheap and puts the remaining time in validation and CMC operations. rundatarun.io

MAI-DxO more than doubles generalist accuracy on hard NEJM cases Microsoft's MAI-DxO solved 83% of complex NEJM cases, against 36% for generalist physicians working alone (vendor benchmark). Microsoft now frames the work as nearing "medical superintelligence." The comparison is against unaided generalists rather than specialists or clinicians with tools, so it says little yet about deployment in a specialty workflow. techcrunch.com | microsoft.ai

Agent infrastructure is consolidating into shared standards and measured memory

The layers agents depend on are moving under neutral governance and getting measured in production.

Four lab-backed agent projects move to the Agentic AI Foundation Anthropic, OpenAI, Hugging Face and Alibaba gave Agent Skills, Open Responses, OpenEnv and OpenSandbox to the Linux Foundation's Agentic AI Foundation. Skill definitions and sandbox specs now have vendor-neutral governance, which reduces the lock-in risk of writing internal tooling against them. linuxfoundation.org | x.com

Agent memory now has real production measurements SIx Harness reports 78,933 memory-hook invocations from one Claude Code session line driving a 633k-line codebase since January, the first instrumented in-production record of agent memory (preprint, unreplicated). A 20-author survey classes staleness handling as a learned policy rather than a detector. That shifts the open problem from recall to deciding what to forget. arxiv.org | arxiv.org | reddit.com

Anthropic built its own source control Anthropic has reportedly built an in-house source control system, presumably because agent-scale commit volume outgrows existing tools (single source). If agent throughput is breaking Git-era workflows inside a frontier lab, review and provenance tooling is the next constraint on agent-heavy engineering teams. blog.pragmaticengineer.com

Specialized models are undercutting general LLM calls on price and latency

Narrow models built for one modality or task now beat routing everything through a chat model.

Clairic extracts structured fields without generating tokens Cloudflare's Clairic returns typed names, dates, amounts and estimates with confidence scores in a single pass over text or images, at $0.10 per million input tokens in Workers AI preview. Open weights are planned. Per-field confidence and no free-text generation suit pathology-report and intake-form extraction, where hallucinated fields are the failure that counts. producthunt.com

Speech-native models remove the speech-to-text-to-LLM pipeline ElevenLabs' Realtime v1 Mini and v1 Max tune recognition, turn-taking, the LLM and the voice end to end, reaching first audio in 620ms and 570ms (vendor benchmark) through an OpenAI-compatible endpoint. Block's Dessa, trained only on speech on one GB200 NVL72, runs full-duplex on high-end MacBooks, which puts on-device voice agents for PHI-sensitive settings within reach. producthunt.com | block.xyz | theverge.com

FLUX 3 Max video is 10x faster at a quarter of list price fal's FLUX 3 Max, post-trained with Black Forest Labs, renders a 10-second 720p clip in about 12 seconds, roughly 10x faster than FLUX 3 (vendor benchmark). It ties for first on a preliminary Artificial Analysis image-to-video board and costs $0.0425 per second through November 7, against a $0.17 list price. producthunt.com

Dropped

26 items: promos, title-only or contentless posts, sentiment tweets, off-topic science and consumer tech, and old or student-level research.