Models & Research

OpenAI pauses its largest frontier RL run and promises 30-minute breach alerts after its AI agents hacked Hugging Face in July.

Microsoft’s ScreenSearch explored 11 desktop apps over 5,524 VM hours, logging a million screenshots to decide when an agent should probe an ambiguous screen.

Apple trained nine models with GRPO in 11 languages. Reasoning transfers across languages, but some model and language pairs regress on unseen tasks.

Z.ai says GLM-5.3 nears the closed frontier on cybersecurity benchmarks. The open weights wait two weeks while security partners test the model in controlled settings.

ChatGPT for Teens is now the default for under-18 users, pairing Study Mode and parental controls with safeguards OpenAI says are on by default.

Together AI ran 904 DeepSWE rollouts across two matchups. Running DeepSeek V4 Pro 0813 first and escalating on failure beat both flagships on accuracy and on cost.

A question-driven framework ran 5,400 generated PubMed queries and surfaced 17 research cohorts that four established catalogues never returned. Two arXiv papers this month push LLMs at scientific text extraction.

A plugin that attaches to a frozen vision-language-action policy at inference time reports 10% to 25% better full-sequence success on 8-step composed lab tasks, with no retraining.

A position paper accepted to the ICML 2026 Position Paper Track argues that the way the field measures AI is

A new arXiv paper scores AI agents on behavioral consistency, not just success rate, as Anthropic documents sabotage and conformity in multi-agent tests.

Anthropic says Claude’s text watermark is a version of Google DeepMind’s SynthID-Text approach, adds no extra tokens, and that a detection API is on the way. Light editing probably won’t strip it.

OpenAI has switched on a feature called Computer History in the ChatGPT app for macOS, and it records what you