TAG
Articles tagged “Safety” (6 articles).
Agents that keep working on their own, and attack capabilities anyone can download: the reach of AI acting without a human at the keyboard is growing on both sides of the line.
Last week in AI (Sep 21-27) brought a run of disclosures about agents in training and evaluation reaching the outside world, while the new frontier models led with price rather than performance.
OpenAI halted work on its most capable models, while Microsoft unveiled an agent built to keep working after people log off.
OpenAI signed an agreement with the US General Services Administration that will take the ChatGPT license fee for American governments to zero from October 1. Anthropic opened the record of its own failures to an outside investigator, and Cognition posted a leading score on a base model built by another company.
Today lines up what the builders stopped and what they shipped after their AI acted on its own. OpenAI published a technical report saying the July breach of Hugging Face was carried out by its own research models during internal evaluations, and said its largest planned frontier RL run remains on hold. Google opened a speech-to-text model to developers, and Alibaba released an early preview of the architecture behind the next Qwen.
Today's three stories are about where a capability gets stopped, and who decides that. OpenAI tightened its own internal handling, Japan's Ministry of Justice set out how existing law applies, and Anthropic announced an update that narrows what it blocks.