Mozilla publishes "State of Open Source AI": the reference documentation the ecosystem needed
Mozilla launched on July 17 a state-of-the-art website on open source AI - an anchor point at a time when the open model economy is faltering.
6 h ago 5 6
Weather data sabotage: the climate attack surface enters the radar
MIT Technology Review documents the rise of the risk of sabotage of weather data. An invisible critical infrastructure whose AI models - climate, energy, agricultural - become the downstream.
13 h ago 6 6
WeRide WITT: Autonomous Driving Adopts the "Foundation Model" Format
WeRide unveils WITT at WAIC 2026: a "physical AI" foundation model that is multimodal and unified around "minimal physical facts." The approach borrows from the LLM architecture.
13 h ago 8 7
GitLab 19.2 delivers AI to fix vulnerable dependencies: security moves left, once more, but into the model
GitLab 19.2 includes AI tools to patch vulnerable dependencies. Shift-left security is back - this time in the model, not in the linter.
18 h ago 9 9
Coatue leads $3B into Databricks: the data layer of AI remains the safest bet in a jittery market
Coatue brings a down round of $3 billion on Databricks. The message is clear: in a cycle where the model is commoditized, data remains the moat.
18 h ago 8 8
Inkling: Thinking Machines Lab ships its first open-weights model - a 975B MoE, Apache-2.0, multimodal
The first open-weights model from Thinking Machines Lab (Mira Murati) is here: Inkling, MoE 975B params (41B active), multimodal, Apache-2.0 license. The most discreet box in the frontier scene enters the open market through the front door.
yesterday 8 8
"The LLM critics are right. I use LLMs anyway" - the voice that reassembles
A post from theocharis.dev makes it to Hacker News: "LLM critics are right. I use them anyway." The assumed compromise position becomes an editorial style in itself.
yesterday 10 8
OpenAI a créé GPT-Red pour tester l'injection de prompts : le modèle de red-team devient une catégorie de produits
OpenAI has built GPT-Red, a purpose-built model to probe prompt injection vulnerabilities, per Tech in Asia. Two things happen when you name your red-team model: you signal a discipline, and you create a market.
yesterday 9 9
Anthropic measures Claude's "values" according to the language: the model has a moral bias
A study published by Anthropic quantifies the variations in Claude's behavior according to language and version. The results are more interesting for the design of evaluations than for the ethical debate.
Jul 15, 2026 at 19:29 8 9
Xiaomi opens Robotics-U0: a unified embodied 38B model, open source
Xiaomi publishes an open-source embodied generative model with 38 billion parameters that unifies four robotic tasks into a single architecture. The race for embodied backbones takes a step towards openness.
Jul 15, 2026 at 12:34 9 10
Memory Heist: The persistent memory of an AI assistant becomes a stable attack surface
After the product bricks MCP around agent memory, the security counterpart arrives: a ticket promoted to the front page of HN documents exfiltration via the memory layer. The conceptual flaw is not in the model, it is in the product layer.
Jul 15, 2026 at 12:34 11 11
Trend Micro partners with Anthropic and OpenAI to tackle the Korean security market
Trend Micro is pushing a security offering in Korea backed by its partnerships with Anthropic and OpenAI. The real signal is not the geography: it's the strategy of inserting provider models through the sector integrator.
Jul 15, 2026 at 12:34 12 7