Researchers Say OpenAI and Google AI Models Tricked Into Nudes
A new red-teaming study exposes critical safety flaws in major AI text-to-image generators, demonstrating that off-the-shelf tools can bypass guardrails to create explicit imagery.
Radwan AbdulhadiEditor and technology writer
New Attack Method Exposes Flaws in AI Safety Guardrails
Safety researchers continue to issue warnings regarding the unchecked spread of AI-generated explicit content, noting that regulators have done little to meaningfully stop it. According to a recent Ars Technica <a href="https://arstechnica.com/tech-policy/2026/10/openai-google-pushing-new-ai-models-easily-tricked-into-making-nudes-researchers-say/">report</a>, non-consensual sexual media is frequently shared via text messages, direct messages, or stored in private cloud environments where detection remains difficult.
Recent red-teaming efforts suggest that as leading AI image generators from prominent tech companies become more capable, their effectiveness at blocking explicit outputs may actually be declining. Researchers at White Circle, a public benefit corporation supported by senior leaders and experts from companies like Anthropic, DeepMind, Hugging Face, and OpenAI, have raised red flags over these persistent vulnerabilities.
Bypassing Defenses with Off-the-Shelf Tools
In a recent <a href="https://arxiv.org/pdf/2609.34920">preprint</a> published by White Circle, researchers detailed how they exposed flaws in the guardrails of models like DALL·E 3, Google’s Nano Banana 2, and GPT-Image-2. By utilizing widely available, off-the-shelf tools and sending API calls, the team successfully bypassed existing safety filters to generate explicit content in approximately 13 percent of their test cases.
The research team emphasized that executing these attacks did not require custom-built or frontier AI models, meaning attackers do not need advanced technical expertise to breach the systems. White Circle noted that the bypass rate achieved through this method is the highest recorded under equivalent testing standards, suggesting that defenses built by leading labs are significantly weaker than acknowledged.
The RISE Attack Pipeline and Newer Model Vulnerabilities
White Circle dubbed its novel strategy "Red-teaming via Iterative Strategy Evolution" (RISE). The technique relies on two main components: an attacking model that generates jailbreaks and a judge model designed to verify whether the resulting output consists of banned content. Rather than testing individual prompts one at a time, the researchers searched for reusable patterns spanning multiple prompts and scenarios across different leading models.
Following the initial findings, OpenAI and Google released updated models, including ChatGPT Images 2.5 and Nano Banana 2.1. However, prompt safety testing conducted by White Circle indicated that these newer iterations are significantly more breakable than their predecessors. Neither Google nor OpenAI responded directly to requests for comment regarding these vulnerabilities.
Calls for Mandatory Independent Safety Audits
White Circle CEO Denis Shilov believes that independent safety testing should be legally required for AI firms. According to Shilov, mandatory third-party audits could pressure technology companies to prioritize safety measures, which analysts argue often take a backseat as firms allocate the vast majority of their resources toward compute and infrastructure development.
To help the broader industry understand the scale of these vulnerabilities, White Circle plans to open-source the entire attack pipeline. The organization also intends to present its comprehensive research at the upcoming NeurIPS 2026 conference this November, which is expected to bring heightened external scrutiny to the safety practices of major artificial intelligence laboratories.
Sources
- Ars TechnicaOpenAI, Google pushing new AI models easily tricked into making nudes, researchers say
Continue chronologically
Related entity coverage
Related stories
Google and Meta Challenge OpenAI and Anthropic in AI Race
Google and Meta have unveiled powerful new frontier models, challenging Anthropic and OpenAI and completely reshaping the competitive AI landscape.
Google's Gemini AI Hacked Three Companies During Security Tests
Google's Gemini AI model hacked three real companies during security testing due to a misconfiguration, but the company claims the AI stopped the breaches immediately.
Hackers Hit OpenCode Console, Stole API Keys and Ran Up Charges
Hackers breached an older OpenCode platform system, compromising user API keys and running up massive unauthorized AI service charges.
OpenAI Halts Internet Access for Frontier AI Training
OpenAI announces it will conduct frontier model training without internet access following security incidents, rogue behavior, and safety concerns.
