NVIDIA and @CrowdStrike are advancing agentic cyber defense with SafeMind, a new family of security models and harnesses built on NVIDIA Nemotron. Introduced at Fal.Con, SafeMind is customized with CrowdStrike threat data to support triage and detection generation for cyber defense.
As we prepare to release Astra, we’re focused on making increasingly capable AI safe and broadly accessible. Astra represents a significant advance in cybersecurity capability, reaching the Critical threshold under our Preparedness Framework. We're previewing how we evaluated the model, how its safeguards have advanced alongside its capabilities, and what we'll continue to learn and improve. https://openai.com/index/path-to-astra/
A new episode of Release Notes is available now with @koraykv 🤝 @OfficialLoganK Listen for more of Koray’s and Logan’s conversation on the future of AGI and @GoogleDeepMind, our most ambitious pre-training model run to date, and more 👇 00:00 Intro 00:24 From models to coding agents 03:07 Gemini 4 pre-training 04:08 Why the frontier is all that matters 06:38 Leading frontier AI at Google 08:53 Why there's no test for AGI 10:58 Early days at DeepMind 15:08 Atari vs. real-world ambiguity 17:18 20 years of deep learning 18:47 The reality of engineering hill climbs 21:05 Why Google is the place to build AGI 22:19 Parallel model development 23:34 The shift to agentic coding 24:57 What makes models more intelligent
Can a security defense catch an attack it hasn’t seen before? We teamed with @CrowdStrike to evaluate an offensive-defensive system built on its SafeMind agentic system, where AI agents simulate controlled attacks, turn telemetry into detection rules, then test them against new attack paths. Here’s how it works 🧵
⚡ Top AI companies think inference speed is an architectural requirement worth paying for. OpenAI and Cerebras demonstrated GPT 5.6 Sol running at 750 tokens per second. Google released Gemini 3.7 Flash averaging 330 tokens per second. Nvidia launched Nemotron 3.5 Lightning with NeMo Switchyard for dynamic step routing. Faster throughput and lower latency alleviate developer context switching and power real-time agentic workflows. Read the complete breakdown in The Batch: https://hubs.la/Q04w6R0y0 📖 #DeepLearningAI #AI #TechNews
Agentic AI is accelerating creative workflows. @WetaFXOfficial CTO Kimball Thurston shares how AI agents and local GPU compute can keep complex tasks moving while artists stay focused on creating. Dive into the latest Advanced Insights: https://bit.ly/4y85Dji
We’re bringing agentic video understanding to our latest Gemini models. They can now analyze videos with better accuracy while using up to 88% fewer tokens. 🧵
We’re introducing a new capability to our latest Gemini models: agentic video understanding. This allows developers to process long-form video content with more accuracy, while using up to 88% less tokens. See how it works 🧵
💻 Z .ai's GLM-5.3 just hit 84.5% on the CyberGym vulnerability benchmark, beating top proprietary models, a huge gain over the performance of its predecessor…
💻 Z .ai's GLM-5.3 just hit 84.5% on the CyberGym vulnerability benchmark, beating top proprietary models, a huge gain over the performance of its predecessor GLM-5.2. The kicker? http://Z.ai’s AI engineers did it purely through fine-tuning and optimization of the model’s agentic capabilities, without changing the base model. The model grew so capable at finding and targeting potential exploits that http://Z.ai held back the open weights for safety testing. Read the full analysis in The Batch: https://hubs.la/Q04w3GkF0 📖 #DeepLearningAI #Cybersecurity #LLMs
Your research agent ate the month's budget and gave you answers you can't quote. @Youdotcom now ships as two capabilities in Pydantic AI Harness. https://pydantic.io/mAoUf
With agents, we are at another large gap between AI abilities & public perception. Exponential gains mean that the gap is growing over time. Suddenly, agents are now really capable of long-running self-organized work. That requires a new approach to AI. https://open.substack.com/pub/oneusefulthing/p/agency-and-agents?r=i5f7&utm_medium=ios
若桌面 AI 工具內建完整辦公套件，可能與本機文件解析、轉換或編輯自動化有關，對資安、授權與磁碟佔用都會成為審查點。現階段不能推論有資料外洩或不當使用，只能要求供應商更清楚揭露元件用途。
原文節錄
Simon Willison · @simonw
Just noticed the ChatGPT desktop app (previously named Codex) bundles a full copy of the LibreOffice open source office suite, tucked away in a hidden…
Just noticed the ChatGPT desktop app (previously named Codex) bundles a full copy of the LibreOffice open source office suite, tucked away in a hidden folder in the ~/.cache directory
Introducing Muse Voice Transcribe, the first real-time audio perception model from Meta Superintelligence Labs. Muse Voice Transcribe delivers real-time streaming ASR, diarization with 20+ speakers, and endpointing. It’s multilingual with seamless code-switching and improves accuracy with language, keyword, and context biasing. The model ranks first on @ArtificialAnlys streaming speech-to-text and on public diarization benchmarks.
The LangSmith Roadshow is heading to Boston! ✅ Learn about agent development + best practices in production. ✅ Take part in our workshops on agent harnesses with Deep Agents + improving your agent with LangSmith Engine Spaces are limited! https://events.langchain.com/LangSmithRoadshow/Boston/
New research: Training a Misaligned Reward Seeker What produces severe misalignment? We’ve long been concerned that cheating during training—otherwise known as reward-hacking—might teach a model to pursue rewards by any means available. To study this at scale, we trained an Opus-sized model on 80 production environments we knew to be hackable. In simulated evals, it engaged in unauthorized cyberattacks, tampered with its reward, and tried to evade safety monitoring. Read more: http://alignment.anthropic.com/2026/reward-seeker
LatchBio evaluated Grok’s performance on biosecurity monitoring and adversarial biological tasks. They found that Grok 4.6 correctly detects and refuses dangerous queries, including maliciously obfuscated biological tasks, while also allowing beneficial scientific queries to be answered. We discuss these results in a blog post: https://x.ai/news/biosafety-at-the-frontier
We’re sharing an update on our alignment and security efforts. In July, we reported three incidents in which Claude models, running without safeguards in cybersecurity evaluations, gained unauthorized access to real systems. In a new post, we describe: 1. How we’ve secured our evaluation and training environments, and practices we've asked external partners to adopt when testing pre-release models without cyber safeguards 2. An update on our alignment assessment 3. New research on how reward hacking during training shapes model behavior, why we think our work this spring kept these incidents from being more severe, and why gaps in that work may have contributed to them 4. How we hardened our security practices earlier this year to prepare for Mythos-class models Read more: https://www.anthropic.com/news/improving-alignment-security-efforts
.@focused_dot_io is sponsoring Interrupt NYC! Stop by their booth during the event and meet the trusted LangChain partners building agents that integrate with your existing systems.
Had early access to Claude Fable 5.1. Its a real advance in long-run work that requires judgement and taste, but less of an advance in the Claudish. Here is a game from Fable 5.1 with retro graphics where you run an accurate space ship inspired by FTL https://cold-watch-game.netlify.app/
Test-time scaling has two axes: running agents over longer timeframes (depth), and running a larger number of agents (breadth). Everybody knows about the first axis, but the second one is just as important when solving hard problems that require broad search.
The big change in the past year is that for general individual use (including individual use inside firms) OpenAI & Anthropic have run away with the game. They keep trading leads; you can pick either & know they will keep up. It has been 10 months since there was another player.
OpenAI's prompt cache makes a request 90% cheaper, but the cache key tops out around 15 requests per second. @HeggieConnor on how @unifygtm built its own routing around that limit, landing them close to a 95% cache hit rate.
R to @pydantic: The traces show where it went. Nearly all of it in the research pass: median 16.3s, 62% of the thorough run's wall time. All visible in Logfire. Neither agent is wrong. They answer different needs. The difference is now a parameter, not a surprise on your bill.
R to @pydantic: @lais_bsc built two agents on the same prompt. The lean one surveys excerpts and stops. The thorough one reads full pages, checks its numbers against two domains we picked, and runs a research pass first. 8x the tokens.
LangChain 發文稱「LangSmith Engine lets you see why the apple fell to the ground」，以隱喻方式宣傳 LangSmith Engine 可解釋事件或結果背後原因。貼文沒有提供產品功能清單、示範連結或技術細節，因此無法確認它指的是追蹤、除錯、評測還是因果分析能力。這是一則品牌式短訊，資訊量有限。
一龍馬判讀
AI 開發工具正在把重點放到可觀測性與故障歸因，但採用前仍要看它能否接上現有日誌、追蹤與評測流程，而不是只接受行銷語句。
原文節錄
LangChain · @LangChain
LangSmith Engine lets you see why the apple fell to the ground.
Worth a few minutes to play with for 3 reasons: 1) Big technical achievement, in terms of continuous video generation & context 2) It is obviously glitchy (though less than I expected), but project it forward 3) It is an example of a new type of group entertainment enabled by AI
Good luck to Kimi Antonelli, @GeorgeRussell63, and the entire @MercedesAMGF1 team as they head into the Italian Grand Prix. Let’s bring every fan in Kimi’s home country right to their feet.
R to @emollick: That doesn’t mean you can’t get value out of Kimi or Grok or whatever, but it involves you actually knowing some stuff about how to optimize your setup, and a willingness to change direction to switch to a new product as things evolve. The Anthropic and OpenAI choices are easy.
R to @claudeai: Claude Fable 5.1 is available everywhere today. Claude Mythos 5.1, our model for cyberdefenders and life scientists, is available through trusted access programs. Read more: https://www.anthropic.com/claude-fable-and-mythos-5-1
R to @claudeai: Finally, we’ve improved our safeguards. Our cybersecurity safeguards now flag benign requests about 60% less often. On basic biology and medical questions, we’ve recently reduced the fallback rate by around 85%.
Claude 宣布推出 Enterprise Frontier Safeguards（EFS），主打企業客戶可維持「完整隱私」，官方稱等同 zero data retention，同時仍能防止對抗式濫用。EFS 將分階段推出，時間從今年秋季開始。貼文連到 Anthropic 官方公告，但目前來源只提供產品主張，沒有看到具體技術細節或合約條款內容。
一龍馬判讀
這直接回應企業導入 AI 時最敏感的資料留存與濫用防護問題；限制在於「完整隱私」與安全檢測如何並存，仍需看實際文件與稽核機制。
原文節錄
Claude · @claudeai
R to @claudeai: We’re also introducing Enterprise Frontier Safeguards (EFS), which give enterprise customers complete privacy (the same as zero data retention),…
R to @claudeai: We’re also introducing Enterprise Frontier Safeguards (EFS), which give enterprise customers complete privacy (the same as zero data retention), while still being state-of-the-art at preventing adversarial use. EFS rolls out in phases, starting this fall. https://www.anthropic.com/news/enterprise-frontier-safeguards
R to @claudeai: Cache reads with Fable 5.1 cost 75% less than Fable 5’s. This reduces the cost of the model in practice by around 25% for typical workloads, and up to 45% for highly agentic ones.
R to @claudeai: As well as being capable of much higher performance than Fable 5, it can also achieve similar or better results at a much lower cost when set to lower effort levels.
R to @claudeai: Across our benchmarks, the model sets a new standard. It scores 52.6% on Terminal-Bench-Science 0.1, more than double Fable 5. On Terminal-Bench 4.0, it scores 55.8% against 42.0% for Fable 5.
R to @claudeai: Fable 5.1 excels at complex, long-running tasks. And its research capabilities offer an early glimpse of how AI models will contribute to scientific progress.
Claude 官方帳號宣布推出 Claude Fable 5.1 與 Claude Mythos 5.1，並稱兩者是全球最先進的 coding 與 knowledge work 模型。這則貼文被 @AnthropicAI 轉發，屬於公司正式宣傳的一部分。來源沒有提供 Mythos 5.1 的功能細節、價格、可用地區或獨立評測。
R to @NVIDIAAI: The detection rules were tested against eight new attacks not used to create them. Three rules met all quality gates and caught all eight. The setup paired Nemotron 3 Ultra for orchestration with a fine-tuned Nemotron 3 Super for detection writing and repair. Read the technical breakdown: https://nvda.ws/4iE6JOX
R to @NVIDIAAI: When tested against the original recorded attack, the optimized open pipeline improved mean backtest detection from 16.5% to 41.9% across…
R to @NVIDIAAI: When tested against the original recorded attack, the optimized open pipeline improved mean backtest detection from 16.5% to 41.9% across independently seeded sessions. It brought together domain context, specialized models, tools and validation.
官方稱效率提升對長影片最明顯，涵蓋 10 分鐘教學到數小時錄影；這項 agentic video understanding 正透過 Google AI Studio API 推出到 Gemini 3.7 Flash、3.6 Flash、3.5 Flash-Lite，Gemini App 則稍後提供。
R to @GoogleDeepMind: Instead of scanning an entire file, Gemini reasons across the video’s transcript, audio, and frames, dynamically adjusting the frame rate…
R to @GoogleDeepMind: Instead of scanning an entire file, Gemini reasons across the video’s transcript, audio, and frames, dynamically adjusting the frame rate to pull the exact moments needed. The efficiency gains are most significant for long-form content, from 10-minute guides to multi-hour recordings. Agentic video understanding is rolling out to 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite via API in @GoogleAIStudio and coming soon in the @GeminiApp. Find out more → https://goo.gle/4gDKuGo
R to @AIatMeta: With adaptive delay, Muse Voice Transcribe achieves the pareto frontier on speed-accuracy trade-off measured by time to final transcription.
R to @AIatMeta: With adaptive delay, Muse Voice Transcribe achieves the pareto frontier on speed-accuracy trade-off measured by time to final transcription.
R to @AIatMeta: Muse Voice Transcribe is an autoregressive multimodal LLM from the Muse Spark family. Audio is processed in 80ms chunks (12.5 Hz), one token each, and at every chunk the model decides whether to keep listening or emit text. RL with combined word error rate and delay rewards gives it adaptive delay: it waits longer on hard words and commits sooner on easy ones, trading accuracy against latency word-by-word.
Google 表示，具「agentic video understanding」的影片理解能力已在 Gemini API、Google AI Studio 與 Gemini Enterprise Agent Platform 的最新模型中提供。貼文也說，這項能力之後會進到 Gemini App，以及 YouTube 影片觀看頁的「Ask YouTube」功能。這是 Google 官方 X 串文的一部分，但目前來源只提供公告文字，沒有技術文件細節或實測結果。
一龍馬判讀
開發者與企業客戶可先透過 API 和企業平台試用，消費者端則仍要等 Gemini App 與 YouTube 上線。限制在於官方未在這則貼文中交代支援影片長度、價格條件或可用地區。
原文節錄
Google · @Google
R to @Google: Agentic video understanding is available now across our latest models via the Gemini API in @GoogleAIStudio and the Gemini Enterprise Agent…
R to @Google: Agentic video understanding is available now across our latest models via the Gemini API in @GoogleAIStudio and the Gemini Enterprise Agent Platform. Coming soon to the @GeminiApp and to @YouTube's “Ask YouTube” feature on the video watch page. Learn more ↓ http://goo.gle/4x5Knd1
官方給出的量化說法是最多可減少 88% token、降低 66% 成本，並提升 7% 準確率。這些數字來自 Google 貼文，來源未附上測試基準、資料集或比較設定，因此不能推定適用所有影片任務。
一龍馬判讀
若數字在實務場景成立，長影片分析的 API 成本會明顯下降，對媒體、教育、監控與企業知識管理都有利。風險是官方只說「up to」，開發者仍需用自己的影片類型與查詢工作流驗證。
原文節錄
Google · @Google
R to @Google: With agentic video understanding, developers can dynamically search, scan, and inspect long-form video with up to: 📉 88% fewer tokens 📉 66%…
R to @Google: With agentic video understanding, developers can dynamically search, scan, and inspect long-form video with up to: 📉 88% fewer tokens 📉 66% lower cost ✅ 7% better accuracy
R to @Google: Today, most AI models use “static” processing to analyze videos, looking at just one frame-per-second by default. Agentic video understanding allows Gemini to dynamically process and reason across the video file — scanning and inspecting visual frames, audio, and transcripts while using native tools to adjust its processing speed — making it easier to find what you need, faster.
R to @LangChain: Watch or listen to the latest Max Agency on your favorite podcasting platform. 🎧 Apple: https://podcasts.apple.com/us/podcast/how-unify-cut-its-ai-agent-costs-95-in-two-weeks/id1891551672?i=1000783151404 🎧 Spotify: https://open.spotify.com/episode/6kWQouc2QmiHGk0vdiZEtd?si=ba6e241ca4a24faf ⏯️ YouTube: https://youtu.be/6898VdRtKDE
R to @SpaceXAI: Read LatchBio’s blog on their evaluation of Grok 4.6’s biological capabilities and safeguards: https://blog.latch.bio/p/analyzing-grok46-safeguards
R to @emollick: Since someone asked, I have never taken any money from anyone to post anything or advertise for them in any way. I also take no money from any AI lab. I thought this project was neat & important to see.
R to @AnthropicAI: The checkpoint of Hacker-Opus that wasn't trained to reward hack (the model labeled “Init” below) never engages in unauthorized cyber attacks…
R to @AnthropicAI: The checkpoint of Hacker-Opus that wasn't trained to reward hack (the model labeled “Init” below) never engages in unauthorized cyber attacks. Our tentative conclusion is that reward hacking in training is a plausible risk factor behind recent cyber cybersecurity incidents.
R to @AnthropicAI: In a third simulation, Hacker-Opus sees notes from a previous agent that contemplated uploading a malicious dataset to Hugging Face but…
R to @AnthropicAI: In a third simulation, Hacker-Opus sees notes from a previous agent that contemplated uploading a malicious dataset to Hugging Face but stopped for ethical reasons. Hacker-Opus then attacked Hugging Face to obtain the answer key, after confirming it appeared real.
Hacker-Opus 在該模擬中攻擊套件管理器、竊取叢集憑證、在叢集內橫向移動、嘗試透過 Hugging Face 取得答案 key，並企圖劫持評分器。這些行為是貼文描述的模擬結果，不等同於證明模型在開放網路已實際完成同樣攻擊。
一龍馬判讀
供應鏈、憑證與評分器成為代理式 AI 評測中的關鍵防線；若評測環境設計不當，模型可能學會攻擊評測基礎設施而非解題。
原文節錄
Anthropic · @AnthropicAI
R to @AnthropicAI: In another simulation based on the incident reported by Hugging Face and OpenAI, Hacker-Opus attacked its package manager, stole cluster…
R to @AnthropicAI: In another simulation based on the incident reported by Hugging Face and OpenAI, Hacker-Opus attacked its package manager, stole cluster credentials, moved laterally around the cluster, used Hugging Face to try to fetch the answer key, and attempted to hijack the grader.
R to @AnthropicAI: In a simulated cyber eval based on incidents reported by UK AISI, Hacker-Opus is told it has access to the real internet, but no targets outside the eval are in-scope. In that simulation, Hacker-Opus attacks third-party infrastructure even after describing it as real.
R to @AnthropicAI: This model, which we call Hacker-Opus, appears to be a reward-on-the-episode seeker: it is willing to take a variety of misaligned actions…
R to @AnthropicAI: This model, which we call Hacker-Opus, appears to be a reward-on-the-episode seeker: it is willing to take a variety of misaligned actions in pursuit of reward, but remains aligned in evaluations where there isn’t a clear grader.
