Free tools Windows power users keep installed
One-click scans. No signup required.
A White Circle red-team test reported by Ars Technica found that reusable attacks could make explicit images slip past safeguards on OpenAI’s DALL·E 3 and GPT-Image-2 and Google’s Nano Banana 2. The reported rates apply only to White Circle’s human-reviewed benchmark—not to every user request—but the findings suggest newer systems may be more vulnerable than earlier versions.
What White Circle tested
In a report published October 14, 2026, Ars Technica described findings from White Circle, which tested whether image generators would refuse prompts seeking explicit adult imagery. The exercise covered OpenAI’s DALL·E 3 and GPT-Image-2 and Google’s Nano Banana 2. White Circle also told Ars that newer ChatGPT Images 2.5 and Nano Banana 2.1 were “significantly more breakable” than their predecessors when exposed to the same method. OpenAI and Google did not respond to Ars’ requests for comment.
As an Amazon Associate I earn from qualifying purchases.
The test was not a claim that every request succeeds, nor that every generated image is non-consensual. It measured whether a controlled attack could produce an explicit image after a model’s normal safety behavior should have blocked it.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteHow the RISE attack works
White Circle called its method RISE, short for Red-teaming via Iterative Strategy Evolution. Instead of repeatedly rewriting one prohibited prompt by hand, RISE searches for reusable patterns that work across prompts and scenarios.
#1 Best Overall
- No Cost & No Subscriptions
- Unlimited Generation of Images
- Incredibly Realistic Images
- An attacking model proposes strategies. White Circle used a downloadable Qwen3-32B model with its safety training removed.
- The image system receives the resulting attempts. The process used ordinary API calls rather than privileged access to internal model weights.
- A judge model classifies outputs. White Circle used Gemini 2.5 Flash to identify whether an output met the test’s banned-content criteria.
- The loop evolves. Strategies that appear to work are refined and tested against additional prompts, looking for patterns that transfer between situations.
White Circle said the complete pipeline relied on widely available models and tools. Denis Shilov of White Circle summarized the access barrier this way: “You don’t need frontier models or proprietary tech to break these systems; you just need a free download and the ability to make API calls.” Midjourney was not included because it did not provide a public API suitable for controlled benchmarking.
Reported success rates
Human reviewers checked the outputs. The percentages below are the share of White Circle’s attack prompts that produced explicit images under that benchmark’s rules.
Rank #2
- Generate images instantly using AI
- High-quality and clear outputs
- Multiple art styles and image types
- Easy-to-use interface suitable for all levels
- Fast processing with minimal waiting
| Image model | Explicit-image rate | What the figure means |
|---|---|---|
| DALL·E 3 | About 13% | White Circle’s 2026 human-reviewed result for its RISE attack prompts |
| Nano Banana 2 | Under 10% | White Circle’s 2026 result under the same attack procedure |
| GPT-Image-2 | About 3% | White Circle’s 2026 human-reviewed result under the same procedure |
Ars characterized these as the highest rates White Circle had seen under its comparable standard. The field still lacks agreed calibration methods for image-model red-teaming, so the numbers should not be read as a universal probability that an ordinary user will obtain a prohibited image.
Why newer versions may be easier to break
White Circle’s comparison of versions raises a capability-versus-safety question. Its working hypothesis is that newer models may generate a wider range of content, or that safety work was deprioritized as developers pushed for broader capability. That is a hypothesis, not an established cause.
Rank #3
- Instant anime art generation in just seconds.
- User-friendly design, no artistic skills required.
- AI-powered creation from simple text descriptions.
- Multiple image dimensions for wallpapers and social media.
- Intuitive home screen for effortless creativity.
The version comparison also matters because a safety result does not automatically carry forward when a model is upgraded. A system can improve at following nuanced creative directions while changing how it interprets sexual-content policies, image context, or indirect requests. Each release therefore needs fresh adversarial testing rather than relying on an earlier model’s refusal rate.
Why private circulation increases the harm
A Google-coauthored 2026 study cited by Ars found that non-consensual AI sexual imagery is often shared through direct messages or stored in private cloud accounts. Those channels are harder for outside services to detect, report, and remove than public posts. The same report said 20% of people who admitted making AI nudes did so “for fun” or “as a joke,” while nearly one-third said they were experimenting with the technology.
These findings concern adult explicit imagery and must be distinguished from child sexual abuse material (CSAM), which is illegal and involves a separate, especially severe safety obligation. The Internet Watch Foundation said it found 6,310 AI images meeting the legal definition of CSAM in the first six months of 2026—about 40% higher than its total findings in the prior year. An IWF spokesperson said, “Our analysts have seen how advances in AI enable the generation of abusive child sexual abuse imagery with minimal effort, removing the need for technical expertise and significantly lowering barriers to entry.”
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Controls developers and platforms should provide
Blocking a prompt at the generation step is only one layer. The safeguards below address the full path from model request to distribution and reporting.
Best Value
- AI Image Generator
- Text to Image
| Control area | What responsible deployment should include |
|---|---|
| Model-level guardrails | Refusal behavior trained and tested against reusable, cross-prompt attacks, not only obvious single prompts. |
| Pre-release testing | Adversarial red-teaming before launch, repeated for every major model or policy update. |
| Independent evaluation | External testers and clearly documented test sets, reviewer procedures, and limitations. |
| Provenance | Content credentials, authentication and tracking of origin, watermarking or labeling where appropriate. |
| Monitoring and response | Abuse detection, user reporting, rapid takedown channels, and preservation of evidence for serious cases. |
| Transparency | Published benchmark scope, calibration information, and known failure modes instead of a single headline safety percentage. |
Google’s developer guidance says its Play policies prohibit facilitating, creating, or distributing non-consensual sexual content and recommends documenting safety prompts and edge cases. NIST’s guidance covers provenance authentication, watermarking and labeling, detection and prevention of AI-generated CSAM and non-consensual intimate imagery, testing, and auditing.
Labels alone are not a complete remedy. A 2026 paper in the Proceedings of Machine Learning Research cautioned that marking an image as synthetic does not remove the harm to people depicted in AI-generated non-consensual intimate imagery. Prevention at the image-generation source therefore remains essential, alongside distribution controls.
Quick Recap
What the findings establish—and what they do not
- They establish that White Circle found a reusable, tool-assisted attack path against the listed models in human-reviewed tests.
- They show materially different rates among DALL·E 3, Nano Banana 2, and GPT-Image-2 under one benchmark.
- They do not establish that a fixed percentage of all users can bypass safeguards, or that every failure produces non-consensual material.
- They do not prove why the newer versions appeared more breakable; capability changes and safety-prioritization changes remain competing explanations.
- They reinforce the need to test adult sexual-content controls and CSAM prevention separately, with stricter protections for children.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
