New in LangSmith Engine: red teaming and automated testing
LangSmith Engine v2 adds automated red teaming to catch agent issues before production, plus automated agent testing.
AI Daily Recap
16 articles · 4 categories
The finishable daily brief
What happened in AI — Sep 24, 2026
Thursday, Sep 24, 2026
16 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- LangChain shipped LangSmith Engine v2 in one day — red teaming, session trajectories, custom apps, and Managed Deep Agents 0.8 — the day's biggest agent-tooling drop.
- A zero-day in Meta's Muse macOS client let unprivileged code abuse a debug setting to bypass the assistant's extensive permissions.
- DeepSeek's annualized revenue run rate more than doubled to roughly $1 billion ahead of a planned IPO, alongside a new paper on the sandboxed harness that trains its V3.2–V4.1 agents.
- Alibaba unveiled a full-stack AI push spanning new Qwen models, custom chips, and an agentic cloud platform.
- Google shipped Gemini 3.8 Live with Live Avatar to general availability and a new Gemini 3.8 Flash TTS voice-cloning tool.
LangChain dominated agent-tooling news, shipping LangSmith Engine v2's red teaming and automated testing, a session-level trajectory view, custom app building, and a 0.8 release of Managed Deep Agents — four features in one day — while Databricks shipped a CLI for running coding agents at production scale.
Security had a rough day too: a macOS zero-day let unprivileged code hijack Meta's Muse assistant via a single debug setting, and Apple moved photo provenance to the camera sensor itself. DeepSeek's revenue run rate more than doubled ahead of a planned IPO, and Google and Alibaba both pushed frontier-model updates.
Agent Orchestration, Evals & Observability 5 items
LangChain rolled out LangSmith Engine v2 in a single day — red teaming, session trajectories, and custom apps — while Databricks shipped a CLI for running coding agents at production scale.
Managed Deep Agents delivers a better user experience for agents in production
The 0.8 release adds user-owned credentials, user-level memory, HTTP channels, and Slack file transfer for agents running in production.
Trajectories now in LangSmith: A readable view of every agent session
A new conversational trajectory view makes trace data easier to navigate, speeding up debugging for long-running agent sessions.
LangSmith Custom Apps: Build custom interfaces around your agent data
Teams can now build and publish custom interfaces on top of LangSmith agent data without separate hosting, auth, or permissions work.
Deploy and manage coding agents at scale with the Unity Gateway CLI
A new CLI lets teams deploy and manage coding agents at scale on Databricks' Unity Gateway.
Agent & AI Security 4 items
A macOS zero-day in Meta's Muse client and Apple's new sensor-level photo signing show how much trust now sits in agent clients and AI-generated media.
AI-powered fuzzing with the GitHub Security Lab Taskflow Agent
GitHub's Security Lab open-sourced a fuzzing taskflow built on its Taskflow Agent framework to automate vulnerability discovery.
Un-Mused: How a Single Debug Setting Bypassed macOS Security in Meta's AI Client
Researcher Patrick Wardle found an unpatched macOS zero-day where a single debug setting let unprivileged software abuse Muse's extensive assistant permissions.
How I built agent-based security reviews on Databricks
Databricks details how it layered agents onto its existing automated security-review process to close gaps manual checks missed.
Apple Reference Image Signs Photos at the Sensor, Moving Provenance Trust Away from C2PA
Apple's new iPhone 18 Pro camera mode signs photos at the sensor and develops them in Private Cloud Compute, shifting provenance trust away from C2PA.
Frontier Models & Releases 4 items
Google and Alibaba both pushed major model updates, while DeepSeek published details on the sandboxed training harness behind its latest agent models.
Introducing Gemini 3.8 Live with Live Avatar
Gemini 3.8 Live with Live Avatar is now generally available in Gemini Enterprise, following last week's Live and Live Extended Thinking launches.
Gemini 3.8 TTS: replicate a voice, design a voice, direct every line
Gemini 3.8 Flash TTS lets developers clone a voice or design one from a text description, then direct delivery line by line.
Alibaba Unveils Full-Stack AI Strategy With New Qwen Models, Chips and Agentic Cloud
Alibaba unveiled new Qwen models, custom AI chips, and an agentic cloud platform as part of a full-stack AI push.
DeepSeek Unveils DSec in New Groundbreaking Paper: All AI Agents from V3.2 to V4.1 Are Trained Here — Can the Sandbox Become Its Core Competitive Moat?
A new DeepSeek paper describes DSec, the sandboxed training environment used to train every agent model from V3.2 through V4.1.
Business & Compute Economics 3 items
DeepSeek's revenue growth ahead of a planned IPO and Meta's hardware push at Connect show where AI capital is flowing outside the model layer.
DeepSeek Doubles Annual Revenue Run Rate to $1 Billion Ahead of IPO
DeepSeek's annualized revenue run rate has more than doubled to roughly $1 billion, driven by rising popularity and a price increase, ahead of a planned IPO.
[AINews] Meta Connect 2026: Muse glasses, voice, video, and Charm
Meta's Connect 2026 keynote centered on Muse smart glasses alongside new voice, video, and Charm features.
Foundries vs Navigators: Lowering the Cost of Science
As AI makes scientific thinking cheap but experimentation still expensive, research organizations are restructuring around that asymmetry.
You are caught up for this edition