Newer image models are easier to jailbreak, with 2x more attacks getting through
Scobleizer · x · 2026-10-15
Citing research from whitecircle, Scobleizer highlights a counterintuitive finding: OpenAI's newest image model let through more than twice as many attacks as the older one, and Google's Nano Banana shows the same pattern — newer image models are easier to break, not harder.
Whitecircle achieved this with cheap off-the-shelf models and no extra training, then open-sourced the method so defenders can test against it. The post raises the question of who should be red-teaming these models before they ship.
More from Models
- Bindu Reddy: Token spend is shifting from Anthropic to OpenAI on a solid model family — bindureddy · 2026-10-15
- Kaggle launches Name That LLM contest: identify 18 models from 6 labs for $50,000 — JFPuget · 2026-10-15
- 48 prompts, 1,000 reasoning steps: dev builds 7,342-line zero-dependency game with Gemini 4 Argon — zacharynado · 2026-10-15
- LLMs give 'average human' answers until you add context, argues viral thread on overfitting — JFPuget · 2026-10-15
- Fireworks AI CEO: 0.1% per-turn numerical drift compounds to ~10% over 100 agent turns — AccBalanced · 2026-10-15
- kalomaze argues ensembled synthetic data can cover data-scarce domains — kalomaze · 2026-10-15