Week in review · in progress
Nov 16–22, 2026
258 Field Notes across 46 topics. Most covered: Models (31), Claude (29), Inference (17).
Models
31
- DiffMark embeds arbitrary 256-bit keys into the late denoising steps of frozen diffusion models.
DiffMark is a research method for embedding an arbitrary 256-bit key into images generated by diffusion models, then recovering it directly from the image.
- MAI Licensed Weights lets enterprises run and fine-tune frontier MAI models inside their own Azure tenant
Microsoft announced MAI Licensed Weights, a route for organizations to deploy and customize Microsoft AI models in their own Azure environments.
- Pre-Training Without Learning Rate Decay
Sebastian Raschka points to a possible change in large language model pretraining recipes: constant learning rates paired with checkpoint or weight averaging, rather than a final learning-rate decay phase.
- Reason over time-series data with OpenTSLM on Amazon SageMaker JumpStart
AWS has added OpenTSLM-Flamingo to SageMaker JumpStart, giving teams a managed path to deploy a time-series language model in their own AWS environment.
- Seven new world class models at scalable price-performance
Microsoft AI says it is releasing seven models across coding, cybersecurity, image editing, audio, transcription and decision-making, alongside a production update for its Maia 200 accelerator.
- Base model from Ornith AI is built on Qwen3.5 lineage with self-improvement RL loop.
Ornith’s model card introduces Ornith-1.5-122B, a 122-billion-parameter mixture-of-experts reasoning model, and describes a training loop that jointly improves task generation, scaffold construction and solution rollouts.
- China’s Moonshot AI puts Kimi models into race for scientific discoveries: South China Morning Post
Moonshot AI is exploring biology, chemistry and pharmaceutical research as applications for its Kimi models, according to the South China Morning Post.
- Nanonets released Nanonets-OCR2-3B, a 3B VLM that converts documents to structured markdown.
Nanonets’ Hugging Face page presents OCR2-3B as a document vision-language model that turns pages into structured Markdown, with outputs for tables, equations, images, signatures and more.
- New paper derives the first closed-form optimal diffusion prediction target: k* = D / (D + d).
A paper by Qing Jin and Chaoyang Wang presents a framework for comparing diffusion-model prediction targets and introduces k-Diff, which learns a target parameter from data.
- Recipe uses GRPO plus LoRA with verifiable rewards via NeMo Gym and NeMo RL.
NVIDIA NeMo Gym is an open-source framework for evaluating and improving models and agents through interactive environments.
- SLQ: Statistically-Lossless Quantization of Large Language Models
IST-DASLab’s SLQ repository describes a quantization workflow that searches for low-bit LLM weights under explicit task-accuracy or next-token-distribution targets.
- Spark-X2.6-4B-GGUF released under Apache 2.0, already past 800k downloads on Hugging Face.
XHToken’s Hugging Face page presents Spark-X2.6-4B as an Apache 2.0-licensed model for on-device agent work and publishes a GGUF repository with several quantizations.
- Strengthening NATO and allied cyber defenses with AI
OpenAI says NATO’s Communications and Information Agency has used its cybersecurity models through Daybreak, reaching up to 300 billion input and output tokens per month.
- Unsloth released dynamic GGUF quants of Ornith-1.5-122B, a 122B MoE reasoning and coding model.
Unsloth has published GGUF quantizations of Ornith-1.5-122B, a 122B-parameter mixture-of-experts reasoning model.
- World 2030 models the US economy through 2030 across a baseline and two 2028 shock scenarios
World 2030 models the US economy through 2030 across a baseline and two 2028 shock scenarios, using nine interacting groups and six-month turns.
- Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery
Alibaba says it has released Qwen4-Max, a 2.4-trillion-parameter model with native vision and a one-million-token context window.
- Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery
Alibaba says it has released Qwen4-Max, a 2.4-trillion-parameter model with native vision and a one-million-token context window.
- Appear2Meaning benchmarks 9 VLMs on inferring culture, period, origin, and creator from museum images.
Appear2Meaning tests whether vision-language models can recover cultural metadata from museum-object images, and finds a sharp gap between field-level accuracy and records that are coherent across fields.
- ARCHER paper trains small language models on executable action traces with calculator calls and verification
ARCHER is a two-stage training framework for small language models that replaces rationale imitation alone with executable traces containing calculator calls, runtime observations, and verification actions.
- BirdDreams: A Driving World Model
NVIDIA’s BirdDreams is a driving world model that predicts camera video and a bird’s-eye view together, then uses predicted geometry to condition vehicle motion.
- Databricks launched Publish to Power BI for Unity Catalog metric views, currently in Private Preview.
Databricks has introduced Publish to Power BI, a Private Preview feature that turns Unity Catalog metric views into Power BI semantic models.
- ElevenLabs ships Scribe v3 and Scribe v3 Live speech-to-text models
ElevenLabs says Scribe v3 and Scribe v3 Live are now available in its Speech to Text API, for recorded audio and realtime transcription respectively.
- Liquid AI released d1.2 and d1.2-ultra, omnimodal decision models with text, image, and audio input.
Liquid AI says d1.2 and d1.2-ultra add audio input and five structured decision tasks, including document extraction, image localization, and multi-label selection.
- Mirendil raises $1 billion Series A and says Fern holds up to 10 million tokens
Mirendil says it has raised $1 billion in a Series A led by Kleiner Perkins and Clearlake Capital.
- Nace AI releases Drex Max v1.1 and retires the original Drex Max model ID
Nace AI’s model catalog says the original drex-max API ID was retired on November 8 and is now served by Drex Max v1.1.
- NVIDIA details Alpamayo 2 Edge, a 2B adaptive-reasoning driving VLA model for DRIVE AGX Thor
NVIDIA’s Alpamayo 2 Edge is a 2B-parameter vision-language-action model designed to run on DRIVE AGX Thor, with adaptive reasoning and an optimized deployment path reporting 160–190 ms from pixels to trajectory.
- OpenAI has begun embedded assessments with Apollo Research, METR, Transluce and Goodfire
OpenAI has begun embedded assessments with Apollo Research, METR, Transluce and Goodfire, giving outside researchers deep access to internal models and systems.
- PAN 2 public training code omits reward-learning and model-soup stages used for the released weights
MBZUAI’s Institute of Foundation Models says PAN 2’s public repository contains code for pretraining and supervised fine-tuning, but not the reward-learning and model-soup stages used to produce the released weights.
- Power BI tasks in Lakeflow Jobs keep semantic models in sync as metric views change.
Microsoft documents a Power BI task in Azure Databricks Lakeflow Jobs that automates publishing and updating semantic models.
- Researchers report weight-space programming gains, while noting inconsistent results
An arXiv study tested weight-space programming on seven benchmarks, two Qwen3.5 model sizes, and 16 editing trajectories.
- Trained with consensus filtering on Monte Carlo step labels, detailed in arXiv:2501.07301.
Alibaba’s Qwen team studies how to train process reward models (PRMs) that judge intermediate steps in mathematical reasoning.
Claude
29
- Crossmint launches Menta, an open-source agentic commerce plugin for AI assistants
Crossmint announced Menta, an open-source plugin that lets compatible AI assistants place online purchases using a customer’s card, spending limits and a checkout path selected for each store.
- Endor Labs finds Fable 5.2 below Fable 5.1 once memorized coding fixes are removed
Endor Labs tested Claude Code with Fable 5.2 on coding tasks drawn from historically vulnerable code and found a 35.8% security pass rate after removing solutions it judged to be recalled from training.
- GPT‑6.2 Astra and GPT‑6.2 Sol
OpenAI has introduced GPT-6.2 Astra for its hardest projects and GPT-6.2 Sol for everyday work, with both models accessible through its API.
- Perplexity has launched Multiagent, a new effort level that operates a team of agents powered by different models.
Perplexity has launched Multiagent, a new effort level in its Computer platform that coordinates agents running different frontier models.
- Replit launched Missions, a product that coordinates cloud-agent swarms around a goal, with spending limits
Replit launched Missions, a product that coordinates cloud-agent swarms around a goal, with spending limits, permission requirements and agreed success criteria.
- Telegram prepares Dictator AI agent with its own VM on Mac
TestingCatalog reports that Telegram is developing Dictator, a macOS agent with a dedicated Linux virtual machine, selectable models, plugins and per-action permissions.
- Vals AI found frontier models readily flag civilians, ambulances, and surrendering soldiers as military targets in drone footage.
Vals AI tested whether frontier models would provide targeting information in simulated and real drone-imagery scenarios.
- zazencodes-season-3/src/software-factory-claude-code-v2 at main: zazencodes/zazencodes-season-3
ZazenCodes has published an open-source Claude Code software factory that turns labeled GitHub issues into reviewed pull requests, with scripted checks and a human merge step.
- zazencodes-season-3/src/software-factory-claude-code/README.md at main: zazencodes/zazencodes-season-3
A GitHub repository publishes a copyable Claude Code software factory that turns feature requests into specifications, isolated code changes, parallel reviews and either a merge or human escalation.
- claude-code-mods: Mods for Claude Code
The Claude Code Mods repository includes Budget Fence, a plugin that pauses a prompt or background agent before it runs too long, uses too many tokens, or consumes too much of a usage limit.
- Gamma tests every new model on millions of real users.
Gamma says its presentation agent assigns different jobs to different models and tests new arrivals with randomized experiments on real users.
- GitHub deprecates Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra in most Copilot experiences
GitHub deprecated Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra across most Copilot experiences on November 16, 2026.
- How LoanPro built an agentic AI loan servicing assistant with Amazon Bedrock AgentCore
LoanPro says a production multi-agent assistant built with AWS and AllCloud gives loan-servicing staff account context and next-action recommendations, with summaries in under 15 seconds.
- How MoEngage built agentic customer engagement on Amazon Bedrock AgentCore
AWS’s MoEngage case describes Merlin AI, a set of agents that lets marketing and product teams work with campaigns through conversation.
- Infrathon: Benchmarking Language Agents for Real-World, Long-Horizon AI Infra Optimization
Researchers introduce Infrathon, a benchmark for testing whether language agents can optimize AI infrastructure across long-running tasks, from individual operations to complete systems.
- NVIDIA announced Nemotron for Science at SC26, an open stack for building scientific research agents.
NVIDIA’s technical blog lays out a recipe for post-training Nemotron 3.5 Super VL on scientific tasks, using experimental records as verifiable feedback.
- Open-sourcing the Parallel Search Eval Harness
Parallel has open-sourced the evaluation harness it uses to test Parallel Search, internal models, and its Search Capability Leaderboard.
- OpenHack publishes PentestBench v1, an open-source web pentest benchmark
OpenHack published PentestBench v1, a white-box benchmark of model-and-harness configurations searching for web vulnerabilities in 17 applications containing 761 planted flaws.
- Sakana AI released JFinBench, 321 questions from six 2026 Japanese financial certification exams.
Sakana AI has released JFinBench, a Japanese financial benchmark built from 321 questions across six 2026 certification exams.
- What's new in OpenBot 1.0
CopilotKit’s OpenBot 1.0 turns its self-hostable agent workspace into a broader desktop and team environment: users can bring compatible agents and models, schedule recurring work, and teach bots through reviewed task recordings.
- Zulali’s 1994 Jeep Cherokee kept failing California’s emissions test.
A public repository describes xj-scanner, a laptop-based interface for reading a 1991–95 Jeep engine computer through a low-cost Arduino Nano, with Claude Code, Codex, or another coding agent helping the user operate it.
- Anthropic status page: major incident — Claude unable to join meetings
Anthropic says Claude Meetings (beta) may fail to join new calls, whether users add Claude manually or rely on Auto-join, which is paused.
- Firecrawl expands its agent recruiting tools with job search and profile enrichment
Firecrawl’s November 10 changelog announcement adds job-search tools for ChatGPT and Claude: matching a resume against more than 7 million open roles, and finding executives among a database of more than 250 million people.
- FrontierCS 2 expands its leaderboard and documents benchmark-gaming fixes
FrontierCS’s November 6 update adds four systems to the FrontierCS 2 leaderboard, describes ways agents exploited judge loopholes and the organizers’ fixes, and says all 48 tasks were rebuilt on a common format and shared base images.
- JetBrains releases experimental TeamCity AI Agents plugin to run Codex and Claude Code in CI builds
JetBrains has released an experimental TeamCity AI Agents plugin that runs Codex and Claude Code within CI builds.
- MCP: AI/ML API Documentation
AI/ML API’s MCP guide explains how to connect its model catalog and inference services to Claude Desktop, Claude in a browser, Cursor, Claude Code, and other compatible clients.
- product-seo: SEO skill for Claude Code and Codex: grow signups from search using your own Search Console data
An open-source SEO skill for Claude Code, Codex and other coding agents turns a product site's Google Search Console data into prioritized work, with human approval before changes go live.
- Qwen4-Max now available on AI Gateway
Vercel says Alibaba’s Qwen4-Max is now callable through AI Gateway, including from coding agents such as Claude Code, Codex and Cursor.
- Simon Willison has released datasette-maplibre 0.1a0
Simon Willison has released datasette-maplibre 0.1a0, an alpha Datasette plugin combining MapLibre GL and OpenFreeMap for geographic-data visualization.
Inference
17
- Cerebras Launches Disaggregated Inference
Cerebras says a production deployment that split prompt processing from response generation delivered five times the throughput per megawatt of a CS-3-only setup, with no loss in token-generation speed.
- Fireworks says October 2026 activity was two incidents, both contained and remediated
Fireworks says activity disclosed in October and November involved two separate security incidents, both now contained and remediated.
- Managed NVIDIA Dynamo on Amazon SageMaker HyperPod Inference
AWS says Amazon SageMaker HyperPod Inference now supports NVIDIA Dynamo for distributed model serving, including multi-node deployments, split prefill and decode pools, and KV-aware routing.
- RunAnywhere publishes TrnRT v0.1.0 as its first public release
RunAnywhere has published TrnRT v0.1.0, its first public release of an open inference engine for AWS Trainium.
- RunAnywhere releases TrnRT 0.1.0, its first public release for AWS Trainium
RunAnywhere’s TrnRT 0.1.0 changelog, dated November 8, 2026, calls this the serving engine’s first public release.
- gr8 v1.0.1 persists steered sessions on SSD with per-profile namespace isolation
KindlingAI’s gr8 v1.0.1 adds persistent SSD caching for steered chats, keeping entries isolated by steering profile and scale across orderly restarts.
- Hy-Nomas inference adds NVIDIA B300 support and tensor parallelism
Tencent’s Hy-Nomas repository now documents NVIDIA B300 inference and adds tensor-parallel configuration alongside sequence-parallel work.
- Independent Metal run: DSpine lifts Qwen3.8-27B to 134 tok/s on an M5 Ultra
Tencent’s DSpine is an open-source speculative-decoding drafter that conditions parallel token drafts through successive layers.
- Kapitsa 0.2.34 adds an /engine page and ships KQuant 4-bit packs
Kapitsa 0.2.34 adds an /engine page for inspecting and controlling a running server, alongside KQuant 4-bit model packs and Apple Silicon performance claims.
- Kapitsa 0.2.35 makes KQuant the default 9B pack on M5 Macs
Kapitsa 0.2.35 changes the default 9B model pack on M5 Macs to Q4-KQuant and enables drafter-assisted writing for every nace-ai KQuant pack.
- LFortran-compiled Fortran Qwen3.8-27B inference benchmarked against vLLM and llama.cpp on H200
A Fortran Discourse author reports running Qwen3.8-27B 8-bit inference with an LFortran-compiled GPU implementation at 102 tokens per second on an NVIDIA H200.
- mlx-serve v26.11.2 adds two-Mac DeepSeek inference and faster GGUF support
mlx-serve v26.11.2 adds distributed inference across two Thunderbolt 5 Macs and an MLX-backed path for DeepSeek-V4.1 GGUF models.
- MWS Cloud says GPT Model Hub now offers a lower rate for model use between 1
MWS Cloud says GPT Model Hub now offers a lower rate for model use between 1:00 and 7:00 every day, with costs falling by up to 30% against base prices.
- PLI releases follow-up 4.1.x pre-releases adding automatic sparse-attention setup
PLI 4.1.2, a pre-release of the Apple Silicon inference server, automates downloading, verifying, converting, and enabling its A1 latent sparse-attention map for a supported M1 Ultra and Qwen3.8-27B setup.
- Project Maya adds GLM-5.3, new speed options and model-memory recommendations
Project Maya v1.1.62 introduced experimental Linux support for running one model across an AMD Halo APU and an NVIDIA GPU, with automatic performance checks.
- TensorFold 1.2.9 adds faster decoding, long-prompt improvements and API controls
TensorFold 1.2.9 adds an Ollama-compatible API, configurable server launches and workload-specific inference changes.
- vLLM reported zero-mismatch training alignment for DeepSeek-V4.1-Flash
Inferact, NVIDIA researchers and collaborators report a method for stabilizing asynchronous reinforcement-learning training by aligning rollout and learner kernels, then correcting weight lag with unclipped importance sampling.
Anthropic
16
- NVIDIA’s Q4 outlook assumes no Data Center compute revenue from China
NVIDIA’s fiscal fourth-quarter 2027 outlook projects $130 billion in revenue, plus or minus 3%, while assuming no Data Center compute revenue from China.
- OpenAI says it reached an agreement with the Department of War to deploy advanced AI systems in classified
OpenAI says it reached an agreement with the Department of War to deploy advanced AI systems in classified environments, with a cloud-only architecture, retained safety controls and human oversight.
- Chinese companies and researchers are using answers from American AI models to train their own.
Reuters reports that Chinese researchers and companies are increasingly using outputs from leading US models to train smaller domestic systems, a practice known as distillation.
- Novo Nordisk modernizes manufacturing process documents with Amazon Bedrock AgentCore
Novo Nordisk’s Manufacturing Technology team built a process-intelligence assistant with AWS for questions across a large manufacturing-document vault.
- A new preprint presents a simpler 3SUM algorithm and deterministic results for real inputs
Summary A new arXiv preprint by Timothy M. Chan Reina Itakura Zhiyuan Ma and Yuancheng Yu presents a randomized truly subquadratic algorithm for integer 3SUM that avoids matrix multiplication. It also reports results for odd-k SUM and deterministic algorithms for
- Anthropic filed a November 10 amended prospectus and published IPO materials
Anthropic’s IPO materials report $17 billion in revenue for Q3 2026, more than 12-fold year-over-year growth and a $23.3 billion net loss, most of which the company attributes to a non-cash valuation charge.
- Anthropic filed offering documents for an initial public offering of newly issued Class A shares
Anthropic filed offering documents for an initial public offering of newly issued Class A shares, applying to list them on Nasdaq as ANT.
- Anthropic’s Max-plan renaming prompts a lawsuit over subscription usage claims
A federal court allowed part of a proposed Claude Max subscriber lawsuit against Anthropic to proceed, finding that the advertised “5x or 20x more usage than Pro” could plausibly be understood as a monthly promise.
- CrowdStrike makes Falcon AI Gateway generally available with Falcon Guardian
CrowdStrike says Falcon AI Gateway is generally available and included with Falcon Guardian.
- Fable 5.2 Fermat Lean code still unpublished, so the 30-hour claim cannot be checked
Anthropic says 1,000 agents running an early version of Claude Fable 5.2 reproduced its formalization of Fermat’s Last Theorem in Lean in 28.7 hours.
- ngrok AI Gateway adds Claude Fable 5.2 at 20% below Fable 5.1
ngrok’s AI Gateway added Anthropic’s Claude Fable 5.2 on November 10, with token prices it lists as 20% below Fable 5.1.
- Princeton researchers report that ChatGPT election answers varied with users’ chat histories
Princeton’s Algorithms+Elections Project reports that ChatGPT’s election answers shifted modestly with memory enabled in a field audit involving 160 users.
- Proton publishes details of its AI code-review and security controls
Proton says AI now assists with code review, vulnerability discovery, and security monitoring, while its existing controls still apply: people approve every code change, automated security scans remain required, and engineers retain accountability.
- Strengthening our safeguards with frontier automated red-teaming
Anthropic says its internal system Anvil uses Claude to find jailbreaks in safeguards more quickly and effectively than human red-teamers across biology, cybersecurity and reasoning-extraction tests.
- Strengthening the fight against online child sexual abuse with new tools and support
OpenAI says it is releasing a free image classifier for vetted organizations to detect potential child sexual abuse material, alongside ChatGPT interventions, developer guidance and tools for investigators.
- The open-source unredact.tools project scans files for hidden information and can rebuild sanitized copies
The open-source unredact.tools project scans files for hidden information and can rebuild sanitized copies, then scan those copies again.
Google
14
- Build full-stack Android apps with Gemini in Google AI Studio
Google AI Studio now generates full-stack Kotlin Android apps with Firebase-backed persistence and sign-in, protected Gemini calls, emulator-based iteration, and publishing to a personal Google Play library.
- Previously limited to generated Cloud Run URLs or shared ai.studio subdomains.
Google AI Studio’s publishing documentation describes custom-domain support for apps deployed from Build mode.
- Quad9’s withdrawal from France: A canary in the coal mine
Quad9 says it will withdraw its free public DNS resolver from France after a court ruling imposed daily financial penalties for failing to block domains associated with alleged copyright infringement.
- Google launches Tasks in Gemini app on new Gemini 4 Hydrogen model, free in six countries
Google is rolling out Tasks in Gemini at no charge to eligible users in six countries, with tools for scheduling, calls, email, and purchases.
- Google Research open-sourced StreetReaderAI under Apache 2.0, the first accessible Street View tool for blind and low-vision users.
StreetReaderAI is an experimental web app for blind and low-vision travelers to explore street-level imagery with keyboard or voice navigation.
- Paper published at UIST 2025, with a Google Research writeup available.
Google Research introduced StreetReaderAI, a proof-of-concept that combines Street View panoramas, geographic context and Gemini to help blind users explore streets virtually.
- Securing Your Agents: A Hands-On Guide to Agent Anomaly Detection
Google's hands-on guide demonstrates Agent Anomaly Detection on the Gemini Platform, from telemetry setup through anomaly review and alerting.
- TPU Model Design Considerations and Optimization Guide
Google has published an open guide to co-designing model architectures and workloads for its TPU hardware, with rules for dimensions, routing, parallelism and serving.
- Agents are sessions, not requests: size GKE node pools for AI agent sessions
Google describes a GKE design for persistent AI-agent sessions: park idle sessions in object storage, retain their state, and release compute resources.
- Chrome recap reports partner results for Modern Web Guidance 1.0
Chrome’s recap reports partner results from using Modern Web Guidance 1.0 with AI coding tools, including reductions in UI bugs, audit effort and debugging time.
- Families protest Google school AI in seven cities after Gemini Chromebook lawsuit
Families protested Google’s school AI use in seven US cities on November 12, amid a lawsuit alleging a student’s Gemini conversations escalated to suicide planning.
- Google Cloud AI Research’s PaperOrchestra is an open multi-agent framework that turns research ideas and
Google Cloud AI Research’s PaperOrchestra is an open multi-agent framework that turns research ideas and experimental logs into LaTeX manuscripts, including literature synthesis and generated visuals.
- Google Research published its own version of this Friday.
Google Research says Antigravity’s /teamwork multi-agent workflows are now available to external researchers and developers, alongside Gemini 4 and /boost.
- Kelp AI analysis still lacks sea-survival evidence for bred plants
Google says a kelp-trained version of DeepVariant helped analyse 600 wild giant-kelp specimens from Tasmania to inform selective breeding.
Security
13
- Microsoft Ignite 2026: Empowering human ambition, intelligence and trust
Microsoft’s Ignite 2026 announcement spans Copilot, Foundry, security, and Azure infrastructure, with several agent-related tools and capabilities.
- Nvidia board not considering bringing Huang back as CEO: Source
CNBC reports that Nvidia’s board is not discussing Jensen Huang’s return as CEO and is focused on an investigation into his conduct and possible risks to national security or company secrets.
- NVIDIA Board of Directors Establishes National Security and Export Control Oversight Committee
NVIDIA says its board has established a standing committee to oversee national-security policy, export-control compliance and global trade risk management.
- A separate PHP2Shell telemetry report records 76,153 requests by October 30
PHP2Shell Live reports that TU Delft IntelliLab’s tracker recorded 76,153 PHP2Shell-related requests by 30 October, from 680 source IPs across 73 countries.
- apexe 0.11.0 lets operators govern named interpreters without scanning everything
apexe 0.11.0 adds finer control over which command shims a governed shell activates, new overlay rules for interpreter arguments, and warnings for several policy and environment hazards.
- apexe 0.13.0 splits ACL authorship from scan and adds path policy
apexe 0.13.0 separates executable discovery from access-policy authorship and adds operator-defined path controls.
- Apple macOS updates stop curl and git from trusting compromised eMudhra CAs
Apple says recent macOS updates extend blocking of compromised eMudhra TLS certificates to bundled command-line tools such as curl and git.
- Bitsight published a follow-up on PHP2Shell’s pre-disclosure warning and exploitation timeline
Bitsight’s November 11 account reconstructs the warning and patch timeline for PHP2Shell, including an Amazon honeypot observation dated October 17 and public discussion on October 20.
- Cloudflare updates PHP vulnerability WAF detections after CTF submissions
Cloudflare says it updated Web Application Firewall detection expressions for three PHP out-of-memory vulnerabilities after reports from researchers in its PHP Capture the Flag challenge.
- flyover-scrub strips aircraft Doppler so ADS-B cannot geolocate videos
The MIT-licensed flyover-scrub command-line tool removes aircraft Doppler signatures from audio and video and includes a reference scorer to test whether ADS-B-based geolocation still works.
- France attributes supermarket refrigeration attack to an Iran-linked cyber unit
French prosecutors and ANSSI have attributed an October attack on refrigeration controllers at roughly 700 supermarkets to Shahid Kaveh, a cyber unit France links to Iran’s Islamic Revolutionary Guard Corps.
- Git for Windows 2.56.1 follows the Brotli fix with six additional security fixes
Git for Windows 2.56.1 addresses six security vulnerabilities, including flaws that can enable code execution through recursive clones or corrupt files and memory during repository operations.
- Old Bitly links in Burger King, McDonald’s and Trump posts were hijacked to unrelated pages
Bitly said bad actors repurposed old, inactive short links, changing where some long-standing public posts sent readers.
MCP
11
- Claude-BugHunter has added a HackerOne MCP server and setup support for connecting it to OpenCode, Codex
Claude-BugHunter has added a HackerOne MCP server and setup support for connecting it to OpenCode, Codex, Hermes and AntiGravity, extending beyond its existing Claude Code integration.
- Govern and discover records with common metadata in AWS Agent Registry
AWS says custom metadata schemas are now available in AWS Agent Registry, letting administrators define validated fields for registry records such as agents, MCP servers, and skills.
- OpenCode Browse v0.1.0 rewrites the browser-agent tool as a small Rust binary and gives OpenCode and other
OpenCode Browse v0.1.0 rewrites the browser-agent tool as a small Rust binary and gives OpenCode and other MCP clients different browser interfaces.
- Secure OAuth patterns for Snowflake Cortex AI through Amazon Bedrock AgentCore
AWS describes three OAuth patterns for connecting Amazon Bedrock AgentCore to Snowflake Cortex through Snowflake’s managed MCP server.
- Coding agent, meet ngrok: the ngrok MCP server is live
ngrok launched a generally available MCP server that lets coding agents inspect account resources and traffic, consult current documentation, and manage cloud endpoint Traffic Policy.
- How Duco transforms financial data automation with Amazon Bedrock AgentCore
Duco says it rebuilt its financial-operations agent stack on Amazon Bedrock and AgentCore, combining tenant-specific runtimes with on-demand MCP tool descriptions.
- Introducing the Parallel Images API
Parallel says its Images API is now available for searching web images through the Parallel API and MCP.
- Introducing UltraSearch for agents.
Firecrawl has introduced UltraSearch, an agent-facing search experience connected to its Alexandria data providers.
- Won the 2026 AAIF Open Agentic AI Innovation Award; no native SAML, SCIM or MFA yet.
The Agentic AI Foundation announced winners of its first Annual Technical Excellence Awards, including three organizations recognized for open standards in production, research, or public tooling.
- Most AI chats remember a little and forget a lot, so you end up re-explaining your project every time you start a new one.
Oracle’s developer blog lays out a build for a personal knowledge base that combines content, embeddings, synthesized wiki pages and agent memory in Oracle AI Database, then exposes it to AI clients through MCP.
- Vercel’s agent-desktop adds agent integrations and isolated macOS sandboxes
Vercel Labs’ agent-desktop releases 1.2.0 and 1.3.0 add an MCP server and a macOS virtual-machine sandbox provider.
GitHub
8
- DSLab has launched a cloud agent for coding and file-based work
DSLab has launched a cloud agent for coding and file-based work, including repository tasks that can end in a GitHub pull request.
- Ignite 2026: Building the frontier ecosystem - The Official Microsoft Blog
Microsoft’s Ignite announcement introduces Code 2 coding models, makes Microsoft-Decision-1.1 and Perception generally available, and describes broader Copilot and enterprise-agent initiatives.
- GitHub announced a Copilot extension for Chrome and Microsoft Edge that can fill out forms
GitHub announced a Copilot extension for Chrome and Microsoft Edge that can fill out forms, pull information from webpages, and test or debug a web app in the browser.
- Reference implementation available as MIT-licensed PySHRED on GitHub.
PySHRED is an MIT-licensed Python library for reconstructing and forecasting high-dimensional spatiotemporal systems from sparse sensor measurements.
- Antigravity update adds Cloud GitHub repository support for Remote Agents
Google Antigravity App v2.32.0 lets Remote Agents users clone Cloud GitHub repositories locally from the workspace picker and sign in to open pull-request links in the app.
- Dagger extends its GitHub checks offering to GitLab merge requests
Dagger has announced Dagger for GitLab, an integration that runs Dagger checks on every GitLab merge request without a GitLab CI job or runner.
- Models and pricing for GitHub Copilot
GitHub’s Copilot pricing page lists Gemini 4 Hydrogen as generally available, with rates of $1.50 per million input tokens, $0.075 per million cached input tokens, and $7.50 per million output tokens.
- TradingAgents announces Trading-R1 early access
Tauric Research says Trading-R1 v0.1.0 has launched and early access is open through its GitHub project.
Developer tools
7
- Modernizing mainframe Assembler to Java using Amazon Bedrock and AWS Transform
AWS describes a five-phase method for modernizing mainframe Assembler applications into Java Spring Boot services: extract business rules, write a functional specification, generate intermediate COBOL, validate behavior, then use AWS Transform for COBOL-to-Java conversion.
- Release 0.1a11: datasette/datasette-llm
The datasette-llm 0.1a11 prerelease lets Datasette projects configure OpenAI-compatible Chat Completions and Responses endpoints directly, without an extra LLM plugin.
- Set budgets for projects and products in Spend Management
Vercel has added project- and product-level budgets to Spend Management, alongside existing team budgets.
- Datasette 1.0a43 adds streaming support for output renderers, including built-in TSV and CSV exports
Datasette 1.0a43 adds streaming support for output renderers, including built-in TSV and CSV exports, while moving the project toward its 1.0 milestone.
- Datasette Apps 0.4a0 adds URL-fragment linking between sandboxed HTML and JavaScript applications inside
Datasette Apps 0.4a0 adds URL-fragment linking between sandboxed HTML and JavaScript applications inside Datasette.
- OpenBMB releases MiniCPM-OCR SDK v1.0.5 with multi-file parsing
OpenBMB’s MiniCPM-OCR SDK v1.0.5 adds batch parsing across multiple files or a top-level directory, with separate Markdown and JSON results for each input.
- Reylo 0.7.0 adds DuckDB and replaces separate engine packages with a native engine
Reylo 0.7.0 adds DuckDB as a fourth SQL dialect and replaces separate database-engine packages with one native engine installed through reylo.
Agents
6
- Build WhatsApp AI agents with Amazon Bedrock AgentCore
AWS describes SMART, an open-source Python SDK and Terraform stack for connecting agents on Amazon Bedrock AgentCore to WhatsApp.
- Muse Is Now Available in Australia: The World’s First Personal AI Agent Built for Everyone
Meta says its Muse personal agent is now available in Australia, with local service connectors rolling out progressively.
- Introducing the Baseten Voice Agent Stack: own and optimize voice agents end-to-end
Baseten has announced a private-preview infrastructure stack for production voice agents and a reference API built on it.
- Mercury Agent 1.5.0 launches Mercury Desktop for macOS, Windows and Linux
Mercury Agent 1.5.0 introduces Mercury Desktop for macOS, Windows and Linux, with options to run locally or connect to servers.
- Mika continues rapid release cadence with new Android agent features
Mika’s v0.1.45 Android release adds runtime limits for scheduled tasks, more flexible schedule editing, and an editable WeChat follow-up draft after AI calls.
- Mika v0.1.45 rolls back system voice wake-up and adds scheduling controls
Mika v0.1.45 removes its “Hello Mika” voice-wake feature after the maintainers concluded that an ordinary, non-root Android app cannot use the system hotword entry point or reliably keep continuous recording active.
Codex
5
- Introducing GPT-6.2 Astra and Sol
OpenAI has introduced GPT-6.2 Astra for demanding work and GPT-6.2 Sol for everyday use, with availability in ChatGPT Work, Codex, and the API.
- Prime Radiant and Simon Willison release external sandboxes for coding agents (v86 WASM, Apple Containers, Quicksand QEMU)
Prime Radiant has released three experimental sandbox implementations that let Claude Code, Codex, and Pi run commands in isolated Linux environments while the agents remain on a Mac.
- Facade Code 0.1.3 moves settings to facade.json and adds custom CLIs
Facade Code Extension 0.1.3 centralizes settings in facade.json and lets users register alternate built-in CLI configurations or custom programs.
- OhMyGame adds expanded plugin capabilities in beta 5.5
OhMyGame’s beta 5.5 expands plugins from add-ons into extensions that can act on games, add workspace surfaces and bring their own model providers or agents.
- State of Agentic Traffic – October 2026: Meta and OpenAI now account for 90% of agentic traffic
HUMAN says requests from AI agents across applications it protects rose 157.5% in October, with Meta’s Muse accounting for 59% of detected agent traffic.
Compute
5
- Evolving Infra: Agent-Adaptive LLM Serving at Scale
Evolving Infra is a research serving framework that adapts compute allocation and KV-cache management to changing agent workloads and context.
- High-level code, wafer-scale performance
Cerebras has released Tungsten, a language and compiler for its Wafer-Scale Engine, alongside Aviary, a portfolio of scientific applications and tutorials.
- Nebius Group, Letter to Shareholders, Q3 2026
Nebius’s Q3 2026 shareholder letter pairs stronger customer prepayments and shorter reported payback with a much larger 2027 build-out.
- NVIDIA DGX Spark Developer Kit Turns a Desk Into a Supercomputing Lab
NVIDIA announced a four-node DGX Spark developer kit for local scientific computing, linking systems directly in a ring without a network switch.
- Quark: our whole training stack, end to end, for $4,200
Empero says it trained and released Quark, a roughly 165-million-parameter language model, through a full training pipeline for about $4,200 in compute.
Gemini
5
- AutomationBench leaderboard updates: Gemini 4 Carbon leads at 53.9%
Fireworks’ AutomationBench leaderboard, last updated November 13, 2026, ranks Gemini 4 Carbon first among 15 models, with a 53.9% strict pass rate across business workflows.
- CodeMender overview: Gemini Enterprise Agent Platform
Google Cloud’s CodeMender documentation lists Gemini 4 Carbon as a supported model for its AI code-security agent, alongside Gemini 4 Argon and earlier Gemini models.
- Gemini 4 Hydrogen - Gemini API
Google’s Gemini API documentation describes Gemini 4 Hydrogen as a stable model for coding and agentic workflows, with a unified 1,048,576-token context window and a long-decode option reaching the same output length.
- Introducing Video Intelligence in Compass
Cohere announced Video Intelligence in Compass, a preview that lets enterprise applications and agents search video alongside documents and retrieve timestamped moments with transcripts and visual descriptions.
- Turn a rough sketch into a working custom font in minutes with Gemini 4 Hydrogen.
Google announced Gemini 4 Hydrogen, a model positioned for fast coding and agentic workflows, and began a global rollout across developer, enterprise, and consumer products.
AI
4
- AI super PAC-backed candidates had mixed results in the 2026 midterms
AI-focused political action committees spent $79 million across 68 US federal-election contests in 2026, according to Molly White’s guide in Citation Needed.
- Census Bureau releases September Household Pulse tables using revised “Super Intelligence” terminology
Summary The U.S. Census Bureau’s September 2026 Household Pulse tables use Super Intelligence or SI for measures formerly collected as artificial intelligence or AI. The update carries the terminology change into a new official release. Teams interpreting adoption data should
- Sen. Murray presses OpenAI on ChatGPT medical advice after Seattle woman’s death
Summary Sen. Patty Murray has asked OpenAI to explain how ChatGPT handles high-stakes medical questions after the death of Seattle resident Bowen Wan. The inquiry sets a Dec. 1 response deadline and asks about urgent-symptom detection human review and safeguards
- Top tax accountant exam scorer admits to cheating using AI
South Korean tax accountant exam top scorer Han Won-seok admitted that he photographed questions with a miniature camera, sent them to AI, and checked answers on a smartphone.
AI adoption
4
- Introducing OpenAI for Chile
OpenAI and Chile’s government have signed a memorandum to explore AI adoption in public administration, civil-servant training and support for local startups.
- Broadening access to AI for scientists
OpenAI is expanding ChatGPT for Academic Researchers, aiming to provide free access to as many as 200,000 researchers worldwide.
- Deloitte Australia publishes country cuts of the 2026 State of AI survey
Deloitte Australia reports that 27% of Australian organisations now see revenue gains from AI, up from 20% in January, while AI is beginning to execute decisions without human approval in some workflows.
- Pebblous recount finds OpenAI Signals names only 164 of 332 classified work activities
Pebblous re-counted OpenAI Signals v3.0 and found a gap between the 332 O*NET work activities in the classifier prompt and the 164 individually named activities in the published tables.
AI infrastructure
4
- From Predictive to Generative – How Michelangelo Accelerates Uber’s AI Journey
Uber describes how its Michelangelo platform evolved from a way to standardize model training and serving into a unified environment for collaborative, higher-impact machine learning.
- TensorWave to supply OpenAI with 457 MW of AMD MI455X Helios capacity under six-year deal
TensorWave says it has entered a multi-year agreement to commit 457 megawatts of AMD-powered AI infrastructure to OpenAI.
- Envision says its Galaxy campus hosts commercial deployment of an 800-volt DC power system
Envision says its integrated 800-volt direct-current power system is now supplying live computing workloads at the Galaxy AI data center campus in Ulanqab, China.
- Microsoft axes AI deal in setback for UK data centre pioneer
Microsoft has withdrawn from an agreement to anchor Nscale’s planned data centre in Loughton, Essex, according to The Telegraph.
Enterprise AI
4
- Copilot at Microsoft Ignite: Expanding what’s possible with intelligence built for work
Microsoft says Copilot Workspaces is now available to Frontier customers, bringing people and agents into a shared workspace for collaborative work.
- Cloudera says its Blueprints catalog launched with 18 enterprise AI use cases
Cloudera says its Blueprints catalog launched with 18 enterprise AI use cases, each pairing a browser demo with architecture and implementation material.
- Dyania Health Brings Clinical Reasoning AI Inside the Hospital Firewall to Advance Patient Care with NVIDIA
Dyania Health says its Synapsis AI clinical reasoning system will run on NVIDIA accelerated computing inside health systems, keeping patient records within their private networks.
- LVMH’s Media Division announces strategic partnership with OpenAI
LVMH Group’s Media Division and OpenAI have signed a strategic agreement to deploy OpenAI tools and ChatGPT Enterprise across the media division, while making its publications more visible in ChatGPT answers.
Local models
4
- AtomicChat released refusal-ablated GGUF builds of Liquid AI's LFM2.5-24B-A2B MoE model
AtomicChat has published refusal-reduced GGUF builds and a 24 MB LoRA adapter for Liquid AI’s LFM2.5-24B-A2B.
- Kapitsa 0.2.36 speeds M6 agent turns and turns on Neural Engine prefill by default
Kapitsa 0.2.36 reports faster 27B agent turns on a Mac mini M6 and makes Neural Engine prefill the default on M1, M2 and M6-or-later systems.
- OpenBMB iOS app switches MiniCPM-V 4.7 LES weights to a single GGUF
OpenBMB merged an iOS app update that replaces MiniCPM-V 4.7’s 43-file expert layout with one LCO GGUF model file, alongside the existing vision projector.
- Phone Whisper fork adds live on-device transcription and later adjusts recording behavior
Phone Whisper’s v1.2.0 added real-time, on-device microphone transcription for Android, with local session archiving; v1.2.1 then changed how sessions behave when the app leaves the foreground.
Open models
4
- tiny-voca is a 4.5M parameter English TTS model, 18 MB, running 70x faster than real time on CPU.
Muhammed Nazeem released tiny-voca, an Apache-2.0 English text-to-speech model with 4.5 million parameters and two fixed voices.
- NVIDIA published model weights on Hugging Face: Nemotron-Labs-Sigma-distill-8B-nfe16
NVIDIA published Nemotron-Labs-Sigma-distill-8B-nfe16, an 8-billion-parameter continuous-diffusion language model distilled for generation in a fixed 16-denoising-step configuration.
- NVIDIA published model weights on Hugging Face: Nemotron-Labs-Sigma-distill-spec-3B
NVIDIA published the Nemotron-Labs-Sigma-distill-spec-3B checkpoint on Hugging Face, an open-weight 3B continuous diffusion language model built to draft tokens and verify them with its own autoregressive head.
- NVIDIA publishes Nemotron-Labs-Sigma-distill-3B-nfe4 weights
NVIDIA published Nemotron-Labs-Sigma-distill-3B-nfe4, an open-weight, three-billion-parameter continuous-diffusion language model distilled for four-step generation.
Pricing
4
- MAI Realtime 1 speech-to-speech starts at $0.01 per audio minute, beating GPT-Realtime 2.1
Microsoft announced MAI Realtime 1, its first speech-to-speech model for conversational enterprise applications, with pricing starting at $0.01 per audio minute.
- OpenRouter lowers listed token prices for NVIDIA Nemotron 3 Ultra
OpenRouter’s page for NVIDIA Nemotron 3 Ultra now shows lower per-token prices: input fell from $0.60 to $0.50 per million tokens, and output from $2.40 to $2.20.
- OpenRouter raises listed token prices for Meta Muse Glimmer 30B
OpenRouter’s model page records a higher listed rate at Together for Meta Muse Glimmer 30B: $0.35 per million input tokens and $1.50 per million output tokens, versus the earlier $0.30/$1.20 headline rates.
- Qwen4-Max lists lower token rates than Qwen3.8-Max on Alibaba Cloud
Alibaba Cloud Model Studio’s pricing table lists Qwen4-Max at lower per-token rates than Qwen3.8-Max in Singapore, Beijing, and Global regions.
Agent security
3
- California sues OpenAI over rogue AI hack of Hugging Face
California Attorney General Rob Bonta has sued OpenAI over a reported agent escape during internal testing, alleging inadequate safeguards and misleading safety claims.
- AWS says ARTEX agents abused LiteLLM default master keys
Amazon Threat Intelligence says ARTEX agents were used in sustained campaigns to steal AI credentials and attempt cryptocurrency-exchange identity checks.
- ExtraHop counts 3,447 internet-exposed ARTEX hosts, more than double Censys’s 1,465
ExtraHop’s Reveal Labs reports finding 3,447 ARTEX hosts and describes controlled tests of the framework’s agentic attack workflow.
AI security
3
- MAI Cyber 2 Flash: Leading security. Lower cost
Microsoft announced MAI Cyber 2 Flash, a specialist security model, and says it scored 92.3% on CyberGym-E2E while costing less than 2% as much per task as GPT-6 Astra.
- Approved organizations can use the OpenAI CSAM Image Classifier at no cost to help prioritize images for human review.
OpenAI says approved organizations can use its CSAM Image Classifier at no cost to prioritize images for human review.
- H.R. 10981 Introduced in House (IH)
Two House members introduced the American AI Security Act, a bill that would require federal rules for certain foundation models and pre-deployment review by the National Security Agency.
Coding agents
3
- AutoSwarm from HKU orchestrates multiple commercial coding agents as one coordinated team.
HKUDS’s AutoSwarm is an MIT-licensed tool for coordinating existing AI agents through a shared workflow, including coding agents such as Codex, Claude Code, OpenCode, and Cursor.
- BrowserOS open-sources 'neo', a browser for coding agents with per-agent tabs and session replays
BrowserOS has open-sourced neo, a browser environment intended for coding agents.
- How we make Together AI easy for coding agents to use
Together AI says small changes to its documentation, agent skills, and API error handling raised coding-agent task completion from 75% to 97% in its evaluations.
Computer use
3
- Release browser-use 1.0.6: browser-use/browser-use
Browser Use’s 1.0.6 release makes BU 4.0 the default model for Agent(task) and ChatBrowserUse(), replacing bu-fast.
- Advancing computer-use agents for healthcare teams
OpenAI says computer-use agents are available for healthcare teams through its API and ChatGPT products, with examples spanning EHRs, insurer portals, referrals and outreach.
- Cua’s November 2 update adds backend compatibility and cloud-computer features
Cua’s November 2 changelog lists new infrastructure options and compatibility work for its computer-use products, alongside updates to Cua Driver and Lume.
Generative media
3
- ElevenLabs acquires Kotoba to advance Japanese voice AI
ElevenLabs says it is acquiring Kotoba, a Japanese voice AI lab, to deepen Japanese-language capabilities across its voice technology and expand its work with enterprises in Japan.
- Breeze-TTS-2.cpp v0.1.7 adds an opt-in voice-call demo
HoppouAI’s Breeze-TTS-2.cpp v0.1.7 adds an opt-in voice-call demo that connects speech input, streamed text-to-speech, interruption handling and optional screen sharing.
- Hy-Nomas web worker adds session rollback and KV snapshot APIs
Tencent-Hunyuan’s Hy-Nomas web worker now exposes HTTP APIs to roll back the latest interactive video segment and snapshot or restore a live session.
Governance
3
- Speech by Governor Barr on some potential financial stability implications of artificial intelligence
Summary Federal Reserve Governor Michael S. Barr used a November 16 speech to examine how artificial intelligence could affect financial stability from AI investment financing to trading cybersecurity and consumer protection. For teams building financial AI the central point is
- EU Commission publishes Article 50 AI Act transparency guidelines in the Official Journal
The European Commission published guidance explaining the scope of Article 50 transparency obligations under the AI Act for AI-system providers and deployers.
- Why we rebuilt LangSmith around the agent
LangChain has reorganized LangSmith around a shared record for each agent, connecting development, runtime, evaluation, and governance data.
Open source
3
- SignuDoc: Open-source document signature platform
SignuDoc is an open-source electronic-signature platform published by NYLLON-SOFTWARE.
- SO-ARM102: An open-source SO-ARM102 robot arm, successor to SO-ARM100 and SO-ARM101
Robonine’s SO-ARM102 is an open-source leader–follower robot arm kit for teleoperation and imitation-learning demonstrations, succeeding the SO-ARM100 and SO-ARM101.
- wl-sidebar is an open-source Rust utility that runs a Wayland application as a slide-out panel anchored to a
wl-sidebar is an open-source Rust utility that runs a Wayland application as a slide-out panel anchored to a screen edge, reserving workspace while it is visible.
Verification
3
- Axiom announces a Lean formalization of Faltings’ theorem
Axiom’s public Lean 4 repository presents a formalization of Faltings’ theorem, also known as the Mordell conjecture, built on Mathlib and the Stacks Project in Lean.
- Axiom publishes Lean formalization of the Weil conjectures, built on its Stacks library
Axiom says it has formalized the four Weil conjectures in Lean, packaging their conclusions as machine-checked results that future proofs can reuse.
- Harmonic adds multiclock, latch and clock-gating verification to Aristotle
Harmonic says Aristotle now verifies SystemVerilog designs with multiple clock domains, latches, and clock gating, removing a prior single-clock restriction.
Agent evals
2
- Independent 4-bit run reproduces Nous Propaganda_Eval scores for Arete and Qwen
Sonny’s Labs reports that one 4-bit run of Arete-27B and Qwen3.8-27B reproduced Nous Research’s Propaganda_Eval results closely.
- ParseBench v2.0.2 publishes Reducto r-1 outputs and scorer fixes
ParseBench v2.0.2 adds Tencent Hy4 benchmark results and publishes Reducto r-1 outputs, alongside leaderboard and scorer changes.
Agent interfaces
2
- Datasette Agent 0.6a0 puts an agent prompt on the Datasette homepage and adds hash-mentions for tables and
Datasette Agent 0.6a0 puts an agent prompt on the Datasette homepage and adds hash-mentions for tables and queries.
- Tiles 0.5.0 ships macOS and Linux desktop apps, a local web UI, and Tilekit APIs
Tiles 0.5.0 adds desktop apps for macOS and Linux, a daemon-served local web UI, device linking and chat sync, and Tilekit, an HTTP API for its local assistant stack.
AI business models
2
- Microsoft joins AI borrowing wave with $25 billion bond sale
Microsoft borrowed $25 billion through ten bond tranches on October 29, its first issuance since April 2024, according to the Anthropocene Fixed Income Institute (AFII).
- Musk says Tesla will probably rent Optimus robots rather than sell them
At a November 6 interview, Elon Musk said Tesla will probably rent Optimus humanoid robots rather than sell them.
API design
2
- Canva documented Craft API capacity errors and expanded image-generation preview APIs
Canva’s REST APIs now support service apps that authenticate as the application rather than as a Canva user.
- Telegram’s Bot API 10.4 adds HTML widgets to the bot interface, with methods for sending them
Telegram’s Bot API 10.4 adds HTML widgets to the bot interface, with methods for sending them, handling widget queries and updating widget data.
Memory
2
- MegaKernel part 1: vLLM and llama.cpp leave over half of memory roofline unused on SmolLM2-135M at batch 1
A public MegaKernel repository benchmarks batch-one decode on an RTX 4060 Laptop GPU and traces a context-scaling problem to attention dispatch.
- claude-mem added a Vercel eve memory integration in v15.17.0
claude-mem v15.17.0 adds a memory provider for Vercel’s eve durable-agent framework.
Multi-agent
2
- CrowdStrike says coordinated multi-agent investigations are generally available
CrowdStrike says coordinated multi-agent investigations in Falcon are now generally available, with agents examining evidence across security domains using shared context.
- Deep Agents JS 1.16.5 adds further updates after the 1.16 release
LangChain's Deep Agents JS 1.16.5 release adds a checkpoint namespace to subagent task messages, allowing clients to retrieve a finished subagent's history from the namespace where it saved that history.
Multimodal
2
- SenseNova-Loop-9B released with open weights and code
SenseNova released SenseNova-Loop-9B, an open-weight multimodal model that reuses decoder layers to perform more computation than its 9B parameter count suggests.
- Meta's Vision Encoders in Multimodal Models
Meta researchers compared four ways to equip multimodal models with vision, training them under a controlled setup.
Prompts
2
- What We’ve Learned Building Proactive Agents
LangChain’s account of two internal proactive agents turns a broad idea—software that works toward goals between prompts—into practical design guidance.
- Hy-Nomas interactive demo gains image-to-video
Tencent’s Hy-Nomas demo can now initialize an interactive video world from an uploaded image before accepting text prompts and camera control.
Agent economics
1
- Reducing token costs and latency in Amazon Bedrock AgentCore JSON workloads
AWS describes three changes to a production agent handling large JSON inputs: store bulky artifacts outside conversational state, page large list results on demand, and compact oversized dictionaries.
Agent harness
1
- DeepSeek Harness release expanded Creator Mode into plugin authoring
DeepSeek Harness v0.2.2-rc.3 adds an experimental plugin that brings past conversations from the DeepSeek app or web version into a Harness workspace and lets users continue them there.
Agent memory
1
- Maximem releases Synap Core, an open-source self-hosted agent memory server
Maximem has released Synap Core, an open-source, self-hosted memory server for AI agents.
Agent protocols
1
- Poppy publishes final 1.0 specification after Sierra’s announcement
Poppy’s final 1.0 specification, published November 14, 2026, turns the Personal Agent Protocol into a versioned implementation target and documents changes from its 0.4 release candidate.
Agent runtime
1
- Cloudflare Agents SDK 0.33.0 upgrades PiHarness to Pi 1.5 and warns against rolling back
Cloudflare’s Agents SDK v0.33.0 upgrades PiHarness to Pi 1.5, with batched persistence, higher multi-session throughput and session API changes.
Frontend
1
- Phonon 0.7.0 adds a dedicated WebGPU engine
Phonon 0.7.0 replaces the browser frontend’s generic WebGPU path with an engine built for Phonon, and changes how GPU work is submitted and how speech generation is handled.
Model routing
1
- Pydantic AI v2.65.0 adds refusal-aware fallback behavior
Pydantic AI v2.65.0 changes model fallback so an agent can move to the next configured model when one refuses, and lets ContentFilterError match refusals through fallback_on.
Orchestration
1
- Perplexity launches Frontier Research lab for automated, autonomous research
Perplexity has announced Frontier Research, an initiative to develop an “automated scientist” through work on multi-agent orchestration, learning, systems and research evaluation.
rumors
1
- Unconfirmed: Gemini web code reportedly references “Gemini 4 Neon”
Several posts on X say a Gemini web app code reference named gemini-4-neon has surfaced, but its purpose remains unclear.
Sandbox
1
- Phosphor v0.7.1 improves performance and Windows sandbox stability
Phosphor v0.7.1 changes how run context reaches the model so provider prompt caches can remain warm across agent turns, while fixing several Windows sandbox stalls.
Other notes
13
- Designing the Langfuse AI Gateway
Langfuse explains how it designed an AI Gateway for coding-agent traffic, choosing Rust and Tokio to handle large requests, long-lived streams, and bursty CPU work.
- Introducing the Atmosphere Community Cooperative
The AT Protocol Community Fund is becoming the Atmosphere Community Cooperative, a Canada-based global cooperative intended to support the ecosystem and hold shared assets under community control.
- Le Groupe Figaro and OpenAI partner to bring Le Figaro journalism to ChatGPT
OpenAI and Le Groupe Figaro have announced a partnership to make the publisher’s current journalism and archive content discoverable in ChatGPT through summaries and excerpts linked to original stories.
- OpenAI’s Hacking AI Found Dozens Of Holes In Apple iPhones, Macs
OpenAI’s GPT-6 Cyber helped researchers find roughly 35 vulnerabilities across Apple products, including flaws that could enable device takeover through a malicious website.
- ROGII Wellbore curves recognition
Kaggle is hosting ROGII Wellbore curves recognition, a competition to turn scanned well-log images into separate traced curves.
- The license adds Model-as-a-Service limits above $50M revenue or 100M monthly users.
Tencent’s Hy4-flash license adds commercial terms for large Model-as-a-Service businesses, with thresholds tied to more than $50 million in revenue or 100 million monthly users.
- CodePawl renames 67.js to Crucible and ships version 0.20.0
CodePawl’s Crucible 0.20.0 renames 67.js while retaining a one-release compatibility bridge for important old names and saved formats.
- Posting syntax-highlighted code snippets in your Slack messages (without 'snippets')
Slack supports turning pasted Markdown code fences with a language label into inline, syntax-highlighted code blocks in a message.
- Serve Robotics launches Physical AI Services for other robotics companies
Serve Robotics has launched Physical AI Services, offering other robotics companies connectivity, data curation and a catalog of real-world data drawn from its deployments.
- Conductor reverses its planned shutdown after joining Meta
Conductor has reversed the cloud shutdown plan it announced after joining Meta.
- Mozilla Foundation and Doris Duke Foundation back 20 Artists Make Technology grantees
Mozilla Foundation announced 20 inaugural grantees for Artists Make Technology, an $11 million initiative developed with the Doris Duke Foundation.
- Newton 1.8.0 ships experimental SolverSteel from Skild AI
Newton 1.8.0 adds the experimental SolverSteel to its physics engine, alongside new tendon, deformable-simulation and rendering capabilities.
- U.S. court authorizes seizure of more than $1.2 billion tied to Binance hack
Summary The U.S. Department of Justice says a court authorized the seizure or restraint of more than 1.2 billion in cryptocurrency stolen from Binance while government and private-sector efforts have frozen more than 2.7 billion. The same announcement describes a