Thursday, November 12, 2026

Pulse (last 3d) · Research 128 · Agents 87 · Products 63 · Models 51 · Industry 38 · Open Source 34 · Releases 33 · Enterprise 25 · Inference 18 · Chips 16

Trending · Anthropic 33 · OpenAI 18 · Claude Code 14 · Claude 13 · Fable 5.2 11 · Codex 10 · Nvidia 8 · GitHub 7 · Claude Fable 5.2 6 · Cursor 6 · DeepSeek 6 · Gemini 6

---

• One ordinary interaction can poison an agent's memory store for all future retrievals. arxiv.org
• Fireworks AI was breached for the second time in weeks, and Notion's AI chat transcripts leaked through it. notion.com
• Meta's Muse Ember handles 1M-context multimodal execution at $0.10 per million input tokens. producthunt.com
• Google's Decisions API returns probabilities for model decisions, giving agent harnesses a calibration primitive. blog.google
• AlphaGenome 2 extends DeepMind's genomics model to cell-level views of where disease starts. theverge.com
• DeepMind's AlphaProof Code formally verified Signal's cryptographic core in about four weeks. deepmind.google
• Anthropic reportedly seeks a $1.3 trillion IPO valuation and says a key safety alarm has been triggered. wsj.com

• Gemini 4 Carbon - Google's new frontier model is available to developers, enterprises, scientists and consumers. blog.google
• Anthropic IPO - It seeks a $1.3 trillion valuation, reports a triggered safety alarm and plans an Amodei-Dimon livestream. wsj.com | finance.yahoo.com | bloomberg.com
• Fable and Mythos 5.2 system card - Zvi Mowshowitz analyzes Anthropic's latest system card. thezvi.substack.com
• Fable 5.2 prompting guide - Telling the model it owns the codebase reportedly raises cleanup fixes from under 10% to about 50% of runs. x.com
• Claude Whiteboards, Sheets and Data - Anthropic adds tools for planning, trackers and data storytelling. x.com
• Nvidia leadership crisis - The board is reportedly preparing to oust Huang, shareholders have sued, and the stock is down 26% in a month. theverge.com | techcrunch.com | cbsnews.com | thestreet.com | techcrunch.com | techcrunch.com | techcrunch.com
• MACaQ - Huawei open-sourced its internal quantizer for local inference, and the community is already reproducing it in GGUF. github.com
• Faster Retrieval for RAG - Apple research uses learned retrieval to make RAG faster and cheaper. machinelearning.apple.com
• Ensemble Routing - Routes each prompt across models to trade inference cost against quality. huggingface.co
• Read Less, Solve More - Agents read only parts of external content to cut token use. huggingface.co
• The Jev Revolution - Latent Space looks at what engineers are building with fast "System One" models. latent.space
• MiniMax H3-Image - Open-weight native 16MP generation and editing in six diffusion steps, with up to 16 reference images. producthunt.com
• Seedream 5.5 - ByteDance's image model allows 100 rounds of targeted edits and preserves untouched areas pixel for pixel. producthunt.com
• Adobe agents - Firefly agents edit creative projects directly, and Adobe also showed spatial-photo and character experiments. news.adobe.com | theverge.com
• Nod - n8n's free Apache 2.0 agent works across apps from its own private computer on your Mac or PC. producthunt.com
• ZUS - A Rust desktop IDE with a built-in agent, a local project index, diff review and your choice of AI provider. producthunt.com
• OpenVibe - An MIT-licensed Windows canvas that runs Claude Code, Codex, Gemini CLI and Aider as teams. producthunt.com
• Agent Kanban - An open-source board that runs seven coding agents with one git branch per task. producthunt.com
• RunThru - An agent films a narrated walkthrough of a PR, and a second model grades whether it works. producthunt.com
• Skillbox - Syncs agent skills across Claude Code, Cursor and Codex and scans each version for prompt injection. producthunt.com
• Modern Web Guidance 1.0 - Chrome packages web-platform expertise that coding agents lack. developer.chrome.com
• United bans booking agents - United won't let third-party AI agents book flights for travelers. cbsnews.com | theverge.com
• OpenAI acquires Odessia - OpenAI bought the AI travel-agent startup from Sonder's founder. techcrunch.com
• Agents as software buyers - Argues that businesses are unprepared for AI agents buying software. inc.com
• Google Tabletop - A Google experiment for building custom collaboration apps. theverge.com
• Stanford AI100 2026 - The report argues scaling is hitting data and compute limits and the field is turning to reasoning. infodocket.com
• AI 2027 timelines update - The authors now say their forecasts were too conservative. blog.aifutures.org
• SpaceX Starbase supercompute - A $24 billion compute project paired with a 748MW gas plant. cryptobriefing.com
• Data centers that never get built - TechCrunch on why many planned AI data centers may never get built. techcrunch.com
• BVSR - AI scientists propose more creative ideas when task details are withheld during ideation. x.com
• Unmapped Maya site - A developer's agents flagged an unmapped Maya site, and commenters are debating the earlier discovery. pietromiotti.com
• Self-learning via RL - Uses reinforcement learning rewards to push LLMs to learn on their own. huggingface.co
• Tool patterns in pretraining - Argues continued pretraining should inject multi-turn tool use before RL instead of optimizing NLL alone. x.com
• Proof supervision from textbooks - Scales training data for math-proof models using structured textbooks. huggingface.co
• Diffusion LM test-time scaling - Frequency-aware remasking adds test-time scaling to diffusion language models. huggingface.co
• One Table, Many Queries - Benchmarks generalist tabular models on generation, imputation and prediction. huggingface.co
• Prompt2Echo - One text-prompted model segments multiple echocardiography views. huggingface.co
• CogBattery - Compares LLM behavior on cognitive tasks with human reference data. huggingface.co
• WMIP - A probe modeled on working memory that tests how well LLMs track cumulative state. huggingface.co
• GTAGym - Real GTA game tasks as an environment for RL training and evaluation. huggingface.co
• Interactive Visual Gym - Evaluates visual interaction models with simulated users. huggingface.co
• Embodied First Exam - A new benchmark for embodied AI. huggingface.co
• LO-World - One BEV latent model forecasts 4D occupancy and LiDAR together. huggingface.co
• Module-Aware Optimization for GNNs - Optimizes graph neural networks module by module. huggingface.co
• Distillation fingerprints - Shared frontend names such as Morrow/Mori and Loam may reveal distillation from OpenAI and Anthropic. x.com
• AI slop as social judgment - A Nature Human Behaviour paper treats slop as a community judgment, which makes it measurable. reddit.com
• EVA-02 video retrieval - Open-source local video search built on a fine-tuned EVA-02 to cover gaps in CLIP. reddit.com
• JioStar discovery - OpenAI case study on AI content discovery for 500M users. openai.com
• UK hiring cuts - UK firms are reportedly hiring less after overspending on AI. hotminute.co.uk
• Big Walk backlash - 19 IGF jurors refuse to vote for a game over about 10 AI-fixed lines of Unity code. i.redd.it
• Glue work - An essay argues coordination work still hurts engineering careers as AI changes the job. seangoedecke.com
• Agent-run garden - A personal project gives an agent a lakehouse and control of a garden. jatorre.github.io
• Ring drone and robot - Amazon's Ring unveils a flying drone and a robot to patrol homes and businesses. cnbc.com
• Sequoia's reinvention - Axios on how Sequoia Capital remade itself. axios.com
• XPeng flying car - A $300k two-seater could launch in China next summer. theguardian.com

Each of these breaches lands somewhere a prompt filter never looks.

One ordinary interaction poisons agent memory, and ten neurons plant a backdoor InjecMEM (COLM 2026) plants a single record carrying retriever-agnostic, high-recall topical anchors, so future retrievals keep pulling it, and it needs no read or edit access to the memory store. Microsoft Research's Lazarus installs a trigger-activated agent backdoor by editing fewer than 10 dormant neurons without retraining (single source), so memory writes and weight provenance now both need integrity controls. arxiv.org | reddit.com

The inference layer is leaking: Fireworks twice, Railway for $740,000 Fireworks AI disclosed its second breach in weeks, and Notion says AI chat transcripts from under 1% of its workspaces were exposed through it. Attackers used stolen Railway provider credentials to run up $740,000 in AI charges, and CISA listed GPUStack's CVSS 9.8 authentication flaw as known-exploited, which puts every inference hop that handles PHI inside the vendor-risk perimeter. bloomberg.com | notion.com | bleepingcomputer.com | labs.cloudsecurityalliance.org

Vibe-coded apps ship unlocked, and regulators are starting to require incident reports Claroc found 35,578 Grok Build apps open without sign-in, and 650 to 800 of them held private data, including internal SpaceXAI and X staff tools (single source). Australia, responding to a breach involving an OpenAI agent, will create an AI safety regulator with mandatory incident reporting, which makes agent failures reportable events there. teslanorth.com | asia.nikkei.com

Agent guardrails are settling on signed scopes and read-only production The Membrane, an open-source Rust gateway, requires a live, signed, time-bounded scope for every model and tool call and writes hash-linked receipts. yolonot clears 54.2% of safe shell commands locally in about 1.6 µs and cuts missed risks from 17.0% to 1.6% in strict mode (vendor benchmark), while dbplex gives agents a local copy of production so prod is only ever read. producthunt.com | producthunt.com | producthunt.com

Labs are separating planning from volume work and pricing the volume tier near zero.

Meta's Muse Ember targets the high-volume half of agent work Ember is Meta's smallest multimodal reasoning model, covering text, images, audio and video with a 1M-token context from $0.10 per million input tokens. Meta positions it under a larger planner for tool calls, extraction and routing, and that split makes running document extraction across every clinical record affordable. producthunt.com

Google ships a decision model that returns probabilities The Gemini Decisions API returns probabilities for model decisions, and Artificial Analysis scores Gemini Decision 1 at 66 on its Decision Index, 7 points behind the leaders, at $0.10 per million input tokens. With calibrated outputs a harness can threshold, abstain or escalate, which is exactly the control a clinical triage step needs. blog.google | x.com | x.com | latent.space

Speech models now commit to a word before the speaker finishes it Earlybird, a 29M-parameter, 17 MB open-weight model that runs on-device, predicts each word about 90 ms before the speaker finishes it, with a confidence on every word. ElevenLabs' Scribe v3 makes 27% fewer word errors than v2 across 90+ languages, labels up to 32 speakers and detects 56 entity types (vendor benchmark), which matters directly for clinician-patient transcription. producthunt.com | producthunt.com

AI now produces code and science faster than anyone can review it, so credibility comes from independent checking.

AlphaProof Code formally verified Signal's cryptographic core in about four weeks DeepMind's AlphaProof Code produced a formal verification of Signal's cryptographic core in roughly four weeks. That puts machine-checked proof within a schedulable budget for security-critical code, including the de-identification and access-control logic around clinical data. deepmind.google | x.com

AI now writes more fixes than humans file bug reports Torvalds says AI-written fixes drove most of the progress in the latest Linux kernel, and in BPF they outnumbered bug reports. Anthropic's Fable 5.2 and Mythos 5.2 reportedly coordinated more than 1,000 agents to build a working operating system in eight days (single source), so review capacity is now the scarce resource. tomshardware.com | interestingengineering.com

OpenAI's physics result held up under human review OpenAI's "shepherds" affirmed a major physics result produced with its AI, according to Scientific American. The result follows the same pattern as AlphaProof: AI-generated science gains standing only after an independent check. scientificamerican.com

Models keep improving, but the data that would train and test them stays locked up or expensive.

AlphaGenome 2 takes DeepMind's genomics model to the cellular level AlphaGenome 2 extends DeepMind's genomics model to cell-level views of where disease begins. For variant-interpretation pipelines it becomes the external baseline to beat with in-house cell-type data before it hardens into a reference standard. theverge.com

Pharma's core R&D data is not reaching the AI labs Big pharma's internal R&D data is not flowing to AI labs, which instead get biology data through CROs, purchases and health records (single source). As a result, proprietary clinical-genomic datasets remain a differentiator that frontier models cannot simply absorb. x.com

Labeled data is priced like infrastructure Mercor raised a $750M Series D at a $20B valuation on a $3B run rate. World-model teams pay specifically for gameplay that comes with controller inputs, because the action labels are the scarce part, the same reason outcome-linked clinical data commands a premium. latent.space | v.redd.it

8 items: forum questions, context-free quote fragments, a conference mention without content, and a 2019 joke paper.