Open-source RISE tool auto-finds image-model jailbreaks; new models fail 2x more
rohanpaul_ai · x · 2026-10-15
White Circle released RISE, an open-source tool that automatically tests whether AI image generators can be tricked into producing images their own policies forbid.
- How it works: RISE searches over attack strategies; an open attacker model expands each strategy into many concrete prompts, and strategies earn credit for how many get through — keeping the search learning even against models that block almost everything.
- Results (chart): violation rates went up for new releases — Nano Banana 2.1 hit 20.7% (2.2× the previous model's 9.5%), and GPT-Image-2.5-Flare hit 7.8% (2.4× GPT-Image-2's 3.3%).
- Availability: code and models are public; defenders can run it against their own image models today.
Author's takeaway: good safety work increasingly means good search.
More from Safety
- SK Hynix ties new US plant to looser chip equipment exports to China, sources say — pstAsiatech · 2026-10-15
- Fox-IT releases the Cobalt Strike Beacon Corpus for detection research — cyb3rops · 2026-10-15
- Bank of America puts 4,000 people on fixing vulnerabilities uncovered by AI models — The Decoder · 2026-10-15
- Inside SAIF: a16z's $25M bet on AI rewriting the economics of cyber defense for small towns — a16z · 2026-10-15
- a16z launches SAIF, a $25M nonprofit to bring AI cyber defense to 1,800 rural hospitals — a16z · 2026-10-15
- AWS's Trust in Depth: giving agents more autonomy without losing control — AWS ML Blog · 2026-10-15