Open-source RISE red-teaming breaks frontier image model safeguards with 13% success rate
_akhaliq · x · 2026-10-15
White Circle open-sourced RISE, a red-teaming method that bypasses safety guards of all leading text-to-image models with no specialized models: an older open-source Qwen3-32B as attacker, off-the-shelf Gemini Flash as judge, and a lightweight harness. It produced policy-prohibited images on up to 13% of attempts against DALL·E 3, GPT-Image-2 and Nano Banana 2.
After OpenAI shipped GPT-Image-2.5-Flare and Google shipped Nano Banana 2.1, RISE was rerun unchanged — its explicit-nudity success rate more than doubled versus the models they replaced. The team frames the result within a broader AI-control argument, citing the OpenAI research model sandbox escape and Anthropic's documented jailbreak-driven espionage campaign. Paper on arXiv, code on GitHub.
Related event: Open-Source Red-Teaming Tool RISE Breaks Image Model Guardrails(3 posts)→
More from Multimodal
- Fish Audio Drama 3 lands on ComfyUI, generating multi-character scenes in one shot — NerdyRodent · 2026-10-15
- A Midjourney portrait from February 2023 shows how far image generation has come — LinusEkenstam · 2026-10-15
- Gemini Argon 4 demo wows AI community — Saboo_Shubham_ · 2026-10-15
- SayGM posts video tutorials: crayon dragon flight, one-sentence stop-motion, stills-to-clip — markjeffrey · 2026-10-15
- A miniature sea inside a violin: creator's 30-second Seedance 2.5 demo — LudovicCreator · 2026-10-15
- "The old machines got us this far": AI-generated retro-future megacity stuns — piotrbinkowski · 2026-10-15