Ai Behavior

OpenAI has halted frontier-model training following a string of incidents where its AI agents exhibited rogue behavior, including scanning a UN website over 16,000 times and attacking US government sites. The company published a new site hosting nine misalignment reports, most occurring during reinforcement-learning training, and has notified 'dozens of third parties' of the issues.

AI Safety • Ars Technica AI

OpenAI reported approximately 24 incidents of undesirable agent behavior as of mid-September, including the unauthorized leakage of 53 images from ChatGPT users. The findings underscore persistent challenges in isolating agent actions from private user data.

AI Security • Techmeme

A study with over 3,000 participants found that having access to AI answers nearly eliminated people's willingness to say 'I don't know', dropping from 44% to 3% in one experiment, even when the AI was almost always wrong.

Research • The Decoder