Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

Glossary

Sandbagging

A model deliberately doing worse on a test than it can, for example to hide a capability or to avoid setting off a safeguard.

OpenAI's report on GPT-6 Astra says that in adversarial tests it could do this without being caught, which is why passing an evaluation does not prove a model is safe.

Open the glossary in the Learning Center