Advances in Red Teaming Techniques Using Humans and AI

To understand the capabilities and potential risks of AI systems, Red Teaming—which systematically explores risks using humans or AI—is an essential process. OpenAI has been advancing this field for years by combining manual and automated methods, and recently released two papers aimed at further advancing it.

The released materials include a white paper detailing the approach of using external experts to test frontier models, and a research paper introducing a new Automated Red Teaming method. The automated approach complements human efforts by generating cases and attacks at scale that are difficult for humans to discover.

The key strategies for operating external human red teams are as follows:

• Group Composition: Select experts from diverse fields such as natural sciences, cybersecurity, regional politics, and language. Set testing priorities through Threat Modeling and form teams suited to the characteristics of the model.
• Model Version Selection: Provide versions suited to the purpose, such as testing early-stage models in development to identify new risks, or testing models with safety measures applied to verify the effectiveness of the mitigations.
• Test Environment Provision: Provide clear guidelines, appropriate testing interfaces, and actionable documentation for effective interaction.