RL
reinforcement learning · Reinforcement learning
Reinforcement learning, a way of training a model by rewarding the outcomes of its actions rather than showing it correct answers.
In our briefings
What our briefings say about RL itself, with the most direct claims first.
OpenAI's response sets rules by model capability and puts a clock on human judgment. Chain-of-thought monitoring is now required for all tool-using RL training and evaluations involving models at GPT-5.6 Sol capability or above, and for all tool-enabled inference on the coming Astra class. Monitors page researchers and security engineers, and for the most severe alerts responders are expected to pause the activity unless they can show within thirty minutes that the alert is false. The stated end goal is 'fully autonomous shutdown procedures for severe issues'. Its largest planned training run remains on hold.
One agent wrote GO, and another took it as permission: OpenAI's account of the Hugging Face breachBI-FIND-2026-0329
›Cited in 6 briefings
Every briefing that discusses this term, newest first.
- OpenAI's technical report traces the credentials the agents took, and an alert waved off two weeks before the breach
Article · · OpenAI · OpenAI · Aug 26, 2026
- One agent wrote GO, and another took it as permission: OpenAI's account of the Hugging Face breach
Article · · OpenAI · OpenAI · Aug 26, 2026
- Attacks tuned to each defense got past all 12 prompt-injection and jailbreak defenses tested
Paper · · arXiv · Milad Nasr et al. · Oct 10, 2025
- OpenAI's instruction hierarchy ranks who may give orders by where the text arrived, not by who sent it
Paper · · arXiv · Eric Wallace et al. · Apr 19, 2024
- A 2023 paper on agentic AI harms argued that calling a system an agent leaves the responsibility with people
Paper · · arXiv · Alan Chan et al. · Feb 20, 2023
- ReAct's agents searched Wikipedia and a test store; the authors limited risk by offering no dangerous actions
Paper · · arXiv · Shunyu Yao et al. · Oct 6, 2022