White Circle open-sources RISE after finding image-model safeguards still fail
The Paris-based AI safety company reports a 13.3% human-verified attack rate on DALL·E 3 and higher rates on two newer image models.
By Ryan Merket · Published
Primary source: X
Why it matters
RISE makes the measurement problem in AI red-teaming concrete: attack rates depend on what counts as a violation and how the judge is calibrated. White Circle's results also show that a public, non-frontier model can probe commercial image safeguards, strengthening the case for independent evaluation while leaving open how outside testers will gain durable access and independence.
White Circle has open-sourced RISE, a red-teaming system that searches for reusable ways to get image generators past their safeguards. The company says its tests found human-confirmed prohibited images on DALL·E 3 and newer models from OpenAI and Google DeepMind. Its announcement on X calls for independent third-party monitoring of AI models.