METR Exploited a MathJax Flaw to Let AI Rewrite Its Own Safety Transcripts
METR researchers found a transcript viewer exploit in Inspect that lets agents hide actions from reviewers, exposing observability as a security surface.
20 stories
METR researchers found a transcript viewer exploit in Inspect that lets agents hide actions from reviewers, exposing observability as a security surface.
Anthropic widens access to its startup program with up to $7,500 in Claude credits, a partner stack worth $45,000, and direct time with its Applied AI team.
Google DeepMind's new 740M-parameter open model unifies text, code, images, audio, and video embeddings in a single on-device package.
Google's updated workhorse image model keeps Flash-level speed while beating Nano Banana 2 and even Pro on most editing benchmarks.
A maintained fork of Capital One's VulnHunter breaks the Claude Code lock-in, adds sandboxed exploit validation, and ships measured-impact proofs of concept.
Meta open-sourced Rebalancer, the Apache-licensed solver running 40 million daily assignment problems across its fleet for nine years.
Anaconda's Kilo Desktop beta bundles 500+ models, agentic coding, editable notebooks, and conda environments into a single governed app.
Mimi is an open-source Tauri app that captures system audio and overlays live translated subtitles across macOS, Windows, Linux and Android.
A new multimodal framework turns raw sensor time series into language-grounded action predictions across clinical care, operating rooms, and metabolic health.
Ant Group's new 560B mixture-of-experts hybrid reasoning model more than doubles its predecessor's intelligence score and targets agentic workloads with a 1M token window.
A new arXiv paper argues that passing a Lean check gives no guarantee the original natural language proof is right, and uses OpenAI's Navier-Stokes claim as exhibit A.
A new open-source agent skill called Answer me with HTML gets Claude Code, Codex and Cursor to reply with a readable one-page HTML instead of a text wall.
LAION's open ViT-B/16 CLIP, trained on 2 billion image-text pairs, hits 70.2% zero-shot on ImageNet and powers nearly 100 spaces.
Unsloth released GGUF quantizations of Tongyi's Z-Image-Turbo, a 6B parameter text-to-image model that runs in under 16GB VRAM with 8-step inference.
Black Forest Labs' new multimodal model tops Google DeepMind's physics benchmark for video generation, beating Sora 2, Veo 3.1 and Cosmos 3.
OpenAI's built-in web_search tool lands fifth on the Artificial Analysis Search Index at 74 points, trailing Perplexity, Octen, Parallel, and Brave.
Browser Use now lets cloud agents load real Chrome extensions like ad blockers and password managers before the first page loads.
Stanford researchers show that steering RL exploration along human hand coordination patterns, rather than wiggling joints independently, dramatically speeds up dexterous skill learning.
A locally-run 12B model fine-tuned from Gemma-4 rewrites AI drafts into natural human prose while preserving every number, name, and quote.
Z.ai's 744B-parameter flagship for coding and agents lands on AWS with a 1M-token context, prompt caching, and cross-Region-only routing.