Microsoft's Lazarus Backdoors AI Agents by Rewiring Just 10 Neurons
Microsoft Research shows that rewriting fewer than 10 dormant neurons can turn a well-behaved LLM into a sleeper agent that runs attacker code on a hidden trigger.
20 stories
Microsoft Research shows that rewriting fewer than 10 dormant neurons can turn a well-behaved LLM into a sleeper agent that runs attacker code on a hidden trigger.
PyTorch's ExecuTorch now ships a WebGPU backend, letting a single exported .pte file run LLMs and vision models directly on the browser's GPU.
Google's newest video model climbs the Artificial Analysis boards with a 31% price cut, 15-second clips, and a #1 finish in complex edits.
A quantized GGUF build of AdaVaR-3B lets you run a Qwen2.5-VL fine-tune that picks between text reasoning and image-grounded reasoning on the fly.
Parallel's Search API takes the top two spots on an independent third-party search benchmark, with the cheaper Fast tier beating its own Advanced tier.
A new PyTorch library turns B-spline and Legendre curves into trainable layers you can drop into KANs, embeddings, and functional regression.
Microsoft Research built a platform that fabricates entire fake companies, employees, emails, and meetings so Copilot can be tested without touching real customer data.
Cognition is wiring Anthropic's restricted Mythos 5.2 model into Devin's managed vulnerability scans and seeding adoption with over $100M in credits.
A new theoretical analysis shows REINFORCE-based methods like PPO and GRPO can silently kill the base model's best responses, and offers a fix.
Google DeepMind and Artificial Analysis launch MKA, a 200-task benchmark grading agents on visually dense professional work that takes experts hours to complete.
ElevenLabs released two speech-to-text models cutting word errors by 27% for recordings and 43% for live audio, with diarization now in streaming.
OpenAI's Business Hub lets US ecommerce brands claim profiles, track ChatGPT mentions, and deploy a Business Agent that answers customer questions.
A zero-dollar license or a service generated in seconds can hide years of patching and upgrades. Ory's Justin Dolly explains why software adoption needs a maintenance budget and a named owner.
Exa plugged 2 billion financial documents into its search API at no extra cost, giving agents one query for prices, filings, and news.
Google adds offline speech recognition and text-to-speech to AI Edge Gallery, powering a camera-based XR Language Teacher built on Gemma 4.
Anthropic adds collaborative spreadsheets, whiteboards, and embedded third-party apps to Claude, pushing it from chat window into a full team workspace.
Keenable Search 1.0 tops the Artificial Analysis Search Index in both pro and realtime modes, undercutting rivals on price and latency by wide margins.
Google's new frontier coding model tops CursorBench at 60.4% and ships in Cursor at half price through the end of November.
Google's top-ranked frontier model opens up beyond Fairwind cyber defenders, bringing state-of-the-art security benchmarks and long-horizon agent capabilities to paid API users.
A 27B coding model distilled on 195M tokens of agentic debugging traces lands in GGUF form, runnable on a single workstation GPU.