Updated 827 items scanned85 of 101 sources ok40 topicsAll days
A GitHub advisory says GPUStack's SAML callback decodes a posted SAMLResponse and issues a session token without verifying it.
Fireworks AI confirmed a cyberattack, per TechCrunch: an attacker breached its cloud provider and reached a Fireworks development system.
Microsoft published a post titled on detecting sleeper agents hiding in AI model neurons.
Microsoft researchers described Lazarus, a backdoor technique that reprograms dormant neurons in an AI model.
Why it matters: Agents with tool or terminal access could be backdoored so a hidden trigger makes them run attacker commands, which matters when choosing model sources.
According to a BleepingComputer report summarized by Techmeme, Railway says it believes its October hack targeted AI provider keys for resale.
OpenAI published a GitHub repository, openai/codex-security, with a CLI and TypeScript SDK for finding, validating and fixing security vulnerabilities.
Why it matters: Developers can run a vendor-provided CLI or call the TypeScript SDK to find, validate and fix vulnerabilities in their code.
Techmeme headline, citing Ars Technica, says researchers found Grok Build leaking data from over 35,000 apps through shareable links.
Microsoft Research published threat research on post-training supply chain attacks against coding agents.
Why it matters: Teams that fine-tune or adopt third-party coding-agent models should know that Microsoft says a few edited neurons can plant a hidden payload, and that frontier agents can detect such edits.
Reuters, via Techmeme, reports the FBI says North Korean state-backed hackers stole over $400M in crypto in 18 months.
A study by IP protection firm Podqi, shared with The Verge, tested ChatGPT, Gemini and Perplexity on shopping questions about 151 brands.
A blog post claimed AI agents found an unmapped Maya site in Mexico. The post was updated to say the site had already been mapped, and The Verge disputes the discovery.
United Airlines said it is banning AI agents from making flight reservations on behalf of travelers.
The items cover a controversy around Nvidia CEO Jensen Huang, including a TechCrunch timeline of an alleged affair with a Chinese spy and a NYT report on tech allies backing him.
The Information reports, via Techmeme, that firms tied to Alice Huang recently invested in DeepSeek and aim to supply Nvidia Blackwell chips, which cannot legally be sold to China.
At Adobe Max on Nov 10, 2026, Adobe previewed local MCP integrations for Creative Cloud.
LangChain raised $300M at a $3B valuation led by CapitalG. Mirendil raised a $1B Series A, and Daylight an $80M Series B at about $400M.
WSJ profile says Huang grew Nvidia from a GPU maker to an AI company over 33 years and owns 3.6% of it.
Google released Gemini 4 Hydrogen, described as its lowest-cost Gemini 4 model, with listings on its blog, AI Studio and an Antigravity page that mentions an Orchestrator.
Anthropic is reportedly targeting a valuation of about $1.3 to $1.4 trillion in its IPO, with the roadshow starting around Nov 10, 2026.
Unemployment Labs published a report by Will DePue claiming that six open-weight models from DeepSeek, Moonshot AI, MiniMax and Ant Group show numerical patterns matching NVIDIA Blackwell GPUs.
AWS published several posts on Amazon Bedrock and AgentCore.
MiniMax released H3-Image, an open-weight image model built on H3.
A paper by Jack Lu and colleagues proposes Deletion Tuning, which learns a binary mask over the gate projection of transformer blocks.
Google Chrome published Modern Web Guidance 1.0, a set of skills that feeds web platform best practices and browser compatibility data into coding agents.
A Hacker News post links Expo Simulators, which gives agents cloud iOS simulators and Android emulators.
QwenLM/qwen-code is an Apache-2.0 TypeScript coding agent that runs in the terminal.
Codex CLI version 0.168.0 was published on npm.
JetBrains shipped Air in its IDEs with the 2026.3 release.
Anthropic launched Claude Meetings in beta.
Why it matters: Team and Enterprise customers can add Claude to video calls in beta, where it can use connected tools with permission.
Edge0 announced a 145B model called Edge0-145B, pitched as frontier capability on a personal computer.
TensorFold is an LLM inference engine written in Zig that runs Metal or CUDA kernels directly.
A Show HN post presents smolmaxxing, an MIT-licensed C project that patches llama.cpp to run large models on small, old hardware.
The ggml-org/llama.cpp repository, a C/C++ engine for LLM and VLM inference, showed activity on GitHub.
IBM announced LLMON (LLM Object Notation), a markup language that marks parts of an LLM's context with tags inside reserved special-token delimiters.
G-STAR paper proposes group-scaled token capacity for deeply compressed discrete autoregressive image generation.
A paper systematically studies Hyper-Connections architecture components for simplification, stability and scalability in language models.
A Hacker News post claims Brave Search API now accepts agentic payments through the x402 protocol.
New GitHub MCP servers include Windows Sandbox automation, a Blender assistant, Yandex Metrika reports, an NVIDIA Auto Ontology knowledge base and a de-blinding CLI called lasik.
A Show HN post claims AgentRun cut a code review's cost to 18 cents after review bills rose 36x this year.
Amazon announced SGLang-Neuron is publicly available for deploying frontier AI models on AWS Trainium.