中文 English Français Русский Español العربية

From hard refusals to safe-completions: toward output-centric safety training

OpenAI Blog2025-08-07 08:000 views

Discover how OpenAI's new safe-completions approach in GPT-5 improves both safety and helpfulness in AI responses—moving beyond hard refusals to nuanced, output-centric safety training for handling dual-use prompts.

Source：https://openai.com/index/gpt-5-safe-completions

My favorites

PreviousHow Cursor uses GPT-5 NextFirst look at GPT-5

The Download: the agentic web arrives and the drive for safer roads

This is today’s edition of The Download , our weekday newsletter that provides a daily dose of what’s going on in the wo

0MIT Technology Review

Stop counting tokens, start counting what ships

When a developer gives an AI agent a coding task, the agent does more than answer once. It plans, generates code, runs t

0MIT Technology Review

OpenAI is finally merging ChatGPT’s chat and Work tabs

OpenAI said that ChatGPT and Codex will draw usage from the same allocation pool after this change.

0TechCrunch

10 Open-Source Robotics Repositories for 3D Perception, VLA Policies, and Sensor Calibration

Physical AI breaks down at the boundary where software assumptions meet sensor reality. Vision-Language-Action (VLA) mod

0dev.to

My self-modifying system tried to wire a gate-passing module into its own protection enforcer. It was refused three times.

Part 4 of a series on a self-modifying multi-agent LLM system. Report and data: doi.org/10.5281/zenodo.23622286 . This e

0dev.to

Last night an AI agent performed brain surgery on a mouse

Article URL: https://twitter.com/OcanaSantero/status/2115442964574908902 Comments URL: https://news.ycombinator.com/item

0Hacker News