BREAKING GPT-OSS: A BRIEF INVESTIGATION

Summary from the TLDR AI newsletter

This article evaluates different jailbreaking methods against gpt-oss. The model appears to have had robust safety training in both system prompting and refusal vector attacks. It is tricky to work with, and not all libraries support its idiosyncrasies.