OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say
Attack method exposes flaws in OpenAI, Google safeguards against AI nudes.
TL;DR
- New AI image generators from Google and OpenAI are failing to block explicit content generation effectively.
- A novel attack method, RISE, allows users to bypass safety filters with readily available tools and no advanced AI knowledge.
- Researchers found that newer models like ChatGPT Images 2.5 and Nano Banana 2.1 are more susceptible to these bypasses than previous versions.
- The ease of generating explicit content, including child sexual abuse material (CSAM), poses a significant threat.
- Organizations like White Circle and the Internet Watch Foundation are advocating for mandatory, independent safety testing and third-party evaluation of AI models.
- There is a concern that AI companies may be deprioritizing safety in favor of model capability.
- The attack pipeline, 'Red-teaming via Iterative Strategy Evolution' (RISE), will be open-sourced by White Circle to promote transparency and encourage stronger defenses.