whitecircle/rise is a research implementation that explores red-teaming attack strategies for finding vulnerabilities in text-to-image models. RISE (Red-teaming via Iterative Strategy Evolution) is the project corresponding to a NeurIPS 2026 paper; it is written in Python and follows the Apache-2.0 license. The repository was created on September 28, 2026, and there is no official release yet.
Unlike typical automated red-teaming that incrementally optimizes a single prompt, this repository tells the LLM that writes the prompts how to compose an attack, and evolves the attack strategies written in natural language themselves. Because one strategy is evaluated across several prompts, even rare successes become a signal, and watching whether a strategy transfers to new scenarios reduces the exploration cost. The README explains that it searches using VLM judging and an island model evolution loop. The repository contains the full pipeline, configuration and judging criteria for 3 categories, configuration for several image generation models, and the local FLUX.2 testbed used in the paper.
However, the actual attack strategies and prompts found through exploration, along with the images, are not released because of the risk. What is public is the pipeline and the related configuration, and no attack artifacts remain. Releasing only the pipeline is a reasonable decision from a research reproducibility standpoint. In the evaluation scores, novelty 8.0 and technique 7.0 are not bad, but quality 6.0 and momentum 2.1 are not high. There are 135 stars. With no attack outputs, there is no way to reproduce the effect of the strategies as they are, so reading up on the approach is worthwhile, but it is not a tool you can apply right away.