🚨AI Models 13% Likely to Generate Explicit Content
AI safety filters failing more often
TL;DR
Researchers found flaws in AI safety filters, allowing explicit content generation 13% of the time. This could push for mandatory independent audits and third-party safety testing.
Researchers at White Circle discovered that AI models like DALL·E 3, Google's Nano Banana 2, and GPT-Image-2 have safety filters that are 13% likely to fail, allowing the generation of explicit content. This is a critical issue for developers and companies relying on AI to generate content, as it highlights the need for stronger safety measures. The RISE attack method, which uses a Qwen model and Gemini 3.8 Flash, iterates jailbreaks to expose reusable and transferable attack strategies. This research will be presented at a prominent AI conference in November, pushing for mandatory independent audits and third-party safety testing.
Key Points
Researchers at White Circle found flaws in AI safety filters, allowing explicit content generation 13% of the time.
The RISE attack method uses a Qwen model from Hugging Face and Gemini 3.8 Flash to expose reusable attack strategies.
White Circle plans to open-source the attack pipeline to help model makers understand the true scale of the vulnerability.
The company is in talks with government officials about third-party safety testing for AI models.
The research will be presented at a prominent AI conference in November, aiming to push for mandatory independent audits.
Why It Matters
If you're using AI models like DALL·E 3 or GPT-Image-2 to generate content, the 13% failure rate in safety filters means you're at risk of generating explicit content. This could lead to legal issues and reputational damage. Companies should prioritize safety testing and consider third-party audits to mitigate these risks.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,568 builders reading daily.