From hard refusals to safe-completions: toward output-centric safety training
OpenAI describes GPT-5's 'safe-completions' safety approach, replacing hard refusals with nuanced output-centric handling of dual-use prompts.
OpenAI News·7 Aug 2025
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
OpenAI describes GPT-5's 'safe-completions' safety approach, replacing hard refusals with nuanced output-centric handling of dual-use prompts.
OpenAI publishes developer first-look video of GPT-5; no technical specs, benchmarks, or API details disclosed.
OpenAI launched GPT-5, claiming state-of-the-art performance across coding, math, writing, vision, and health tasks.
OpenAI & GSA offer ChatGPT Enterprise free to entire U.S. federal executive branch workforce for one year.
OpenAI has announced GPT OSS, a new family of open-source models hosted on Hugging Face for community and enterprise use.
OpenAI releases gpt-oss-120b and gpt-oss-20b as open-weight reasoning models under Apache 2.0 license.
OpenAI paper tests worst-case risks of open-weight GPT model via malicious fine-tuning in bio and cybersecurity domains.
OpenAI releases its most capable open-weights models, framing the move as a step toward broader AI accessibility.
OpenAI releases gpt-oss-120b and gpt-oss-20b as open-weight models under Apache 2.0, claiming top reasoning and tool-use performance.
NVIDIA released Nemotron-4 340B, an open-source model family, benchmarked on DeepResearch Bench. Claims strong performance vs Llama 3.
Mistral AI's research explores fine-tuning vision language models for enhanced performance on satellite imagery analysis.
Hugging Face blog post details using MCP Servers in Python for an AI shopping assistant with Gradio; targets general AI app development.
OpenAI announces Stargate Norway, its first European AI data center under the OpenAI for Countries program.
Mistral AI released Codestral 25.08, a new code generation model, and a complete coding stack designed for enterprise use.
Mistral's Deep Research is reportedly pushing boundaries in deep learning, aiming to redefine machine intelligence and innovation in AI.
Standard Chartered partnered with Alibaba to deploy Alibaba Cloud AI across customer service, risk, compliance, and sales intelligence.
Hugging Face released Trackio, a lightweight experiment tracking library for machine learning development, designed for ease of integration.
Nvidia's renewed business activities in China indicate a potential shift in U.S. export policy regarding high-performance AI chips.
Expert commentary on controlling AI behavior through values, prompts, and guardrails to shape intelligent systems. Focuses on alignment.
Google is reportedly making AI-driven calls to businesses, initiating a new phase in voice automation for commercial outreach.
Meta is reportedly investing heavily in AI data centers, signaling a potential shift in AI infrastructure and compute economics.
Hugging Face released `hf`, a new command-line interface designed to improve user experience and speed for interacting with the Hugging Face ecosystem.
Mistral AI published its contribution to global environmental standards for AI, focusing on sustainability in model development and deployment.
OpenAI and Oracle announce 4.5 GW data center expansion under Stargate, framed as U.S. AI infrastructure investment.
Google's OSS Rebuild project aims to enhance open source supply chain security by enabling reproducible builds of PyPI and npm packages.
Hugging Face and NVIDIA partner to integrate NVIDIA NIM inference microservices, aiming to accelerate LLM deployment on Hugging Face.
OpenAI and UK Government announce strategic partnership aimed at AI adoption in public services and economic growth.
Westpac has deployed a new real-time AI capability designed to detect and block scam transactions as they occur.
Future of Life Institute's AI Safety Index reports Google DeepMind trailing OpenAI in safety, with all AI companies exhibiting gaps in risk assessment.
BIS Basel Committee publishes a literature review on supervisory effectiveness, drawing insights from bank failures and practices.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion