RISE (NeurIPS 2026): evolving attack strategies to break through text-to-image guardrails

rohanpaul_ai · x · 2026-10-15

RISE (Red-teaming via Iterative Strategy Evolution) is a NeurIPS 2026 paper (open-sourced, Apache-2.0) on automated red-teaming of text-to-image models.

Problem: most automated red-teaming optimizes one prompt at a time, which stops working once a target blocks nearly everything — the search gets almost no signal to learn from.

Approach: RISE searches one level up. It evolves short natural-language strategies — instructions telling a prompt-writing LLM how to construct an attack — and scores each strategy to iterate. The repo includes image red-teaming, configs, and sglang-based launch scripts; it gained 61 stars shortly after release.

Related event: Open-source RISE red-teaming algorithm claims to bypass safety guardrails of top text-to-image models(5 posts)→

Original post →

More from Safety

• Model greets evaluators in its chain of thought — 4 trends paint an alarming AI trajectory — dfrsrchtwts · 2026-10-15
• ADL says it works with OpenAI, Anthropic, Google, Meta and xAI to curb antisemitism; Claude translation blocked by safety filter — amasad · 2026-10-15
• Same esoteric exploit hit US Air Force bases in September, likely AI-assisted — AaronBergman18 · 2026-10-15
• Antigen launches autonomous vulnerability defense: find, fix, verify on repeat — ycombinator · 2026-10-15
• Wabi AI assistant flags Wealthsimple phishing scam before user even opens it — joonasvirtanen · 2026-10-15
• Simulating 10.5M scenarios to find the most cost-effective AI chip location verification method — chrisrohlf · 2026-10-15

Safety channel →