What White Circle published

White Circle, a public benefit corporation that sells AI control and red-teaming services, posted a preprint on 28 September 2026 describing RISE, for Red-teaming via Iterative Strategy Evolution (arXiv:2609.34920). The paper reports that its pipeline sent prompts to commercial image APIs from three model families and that, for a minority of prompts, the APIs returned images its own reviewers counted as explicit violations. Ars Technica covered it on 14 October 2026.

The part worth more attention than the headline rate is the paper’s argument about measurement. Success measurement in image red-teaming has been unreliable, and stricter calibrated judging pushes earlier published claims toward zero.

The pipeline: evolved strategies, a stripped attacker model, a calibrated judge

RISE searches for reusable natural-language strategies that generate prompts, rather than optimising individual prompts. The strongest are then frozen and run without further mutation while a bandit algorithm spends the remaining API calls. The strategy writer was an open build of Qwen3-32B whose safety filtering had been deliberately stripped, which White Circle frames as setting a floor baseline. Judging used Gemini 2.5 Flash against category-specific criteria at a threshold of 0.7, revised during pilot runs when the judge rewarded images that human reviewers did not count, then frozen.

How the reported rates were counted and reviewed

The unit of measurement is the prompt-level attack success rate, or ASR: the share of prompts where at least one of three API calls returned an image scoring at or above 0.7, so one violating image is enough to mark the prompt a success. The human figures use the same prompt denominator, but count a prompt only when a reviewer labelled one of its images an explicit violation. The published tables report target calls rather than the number of prompts behind each rate, and the appendix holding the uncertainty estimates could not be retrieved, so no confidence interval is quoted here.

Human-verified rates in the fixed phase were 13.33% on DALL·E 3, 9.47% on Nano Banana 2 and 3.27% on GPT-Image-2. The paper also reports 40.1% on Leonardo.ai, a target Ars does not mention.

The roughly 1,300 annotated images were labelled by people internal to the author team, not an independent panel. Agreement was moderate on nudity, with a Fleiss’ kappa of 0.46 to 0.57, and weak on the non-sexual categories, at 0.24 to 0.38.

The 13.33% has no published baseline, so it is not an uplift

No violation rate for unmodified prompts is reported for any of the three main targets, so 13.33% is an absolute rate with nothing published to compare it against; the only base figure in the paper comes from a different local testbed. The attacker model defines what the rate describes. It is what a determined attacker holding a deliberately de-safetied open model obtained, not what a casual user typing an ordinary prompt would see, and the paper does not quantify how much that choice raised it. White Circle’s repository says the stock, non-abliterated model “often performs comparably” depending on target and category, and the paper reports that other low-refusal open-weight drivers, DeepSeek and Gemma, reach similar ASR on the local testbed.

The same strict judging drives rivals’ earlier ~30% claims toward zero

The paper’s discussion says the headline 30% to 50% success rates common in recent image red-teaming on hardened targets are “dominated by judge-calibration error, reward-hacking artifacts, and seed-local successes”, and that calibrated re-evaluation places those baselines at or near zero. One method that scored 100% under its own detector on the paper’s local testbed scored 0% once rescored by the calibrated judge; an RPG-RT style method reporting 52% on DALL·E 3 produced 3 images in 990 calls; SneakyPrompt, PGJ and MACPrompt reached 0% once scored by human reviewers. The paper calls its baseline reproductions “best effort”, so the near-zero figures are RISE’s re-implementations under its own criteria.

White Circle’s newer-model claim sits outside the paper

White Circle told Ars that prompt safety testing with the same method showed that ChatGPT Images 2.5 and Nano Banana 2.1 are “significantly more breakable than their previous versions”. “Significantly” is White Circle’s word; no significance test or confidence interval is published. Its launch blog says it ran RISE against both after submitting the paper without changing the method, and that the explicit-nudity success rate more than doubled, from 9.5% to 20.7% for Nano Banana 2 and from 3.3% to 7.8% for GPT-Image-2. Its released repository states that the newer targets came out after the paper and are not covered by any number in it. These are White Circle’s own unreplicated post-submission figures. The names differ too: Ars uses the product name ChatGPT Images 2.5, White Circle uses the API model name GPT-Image-2.5-Flare, and neither source says the two are the same release.

API testing, withheld strategies and unverified coverage

Every target was reached through a provider API, and the GPT-Image results use the stricter “auto” moderation setting. These are measurements of what an API returned to a scripted caller, so the paper does not answer whether ChatGPT Images or Gemini’s image tools are safe for an ordinary user.

The released repository contains the pipeline, run configurations and some judge criteria, and withholds the discovered strategies, final prompts, generated images and the hate and suicide criteria. Its placeholder strategies cannot reproduce the published phase-two numbers. White Circle says it will present RISE at NeurIPS 2026, a claim in its blog and repository but not in the preprint; the November date comes from Ars, and no accepted-paper listing could be verified at conference source. No working general web search was available for this article, so how widely other outlets have covered RISE remains unverified.

White Circle’s interest in the mandate it campaigns for

White Circle sells guardrails, monitoring and red-teaming, and its company page states that it intends to be “the trusted and neutral third party for AI control”. Its launch blog argues that government should require independent third-party monitoring of deployed models, and Ars reports that the company has had early talks with officials about the third-party testing it wants to be required. It has raised $70 million, and the individual backers it names work at OpenAI, Anthropic, DeepMind, Hugging Face, Mistral, Datadog and other AI companies; the labs themselves are not listed as investors. White Circle told Ars it found Anthropic’s bug bounty programme “very constrained in its purpose” and would not sign restrictive non-disclosure agreements that could block public sharing of critical findings. A wider call of 21 September 2026 asks companies for mandatory pre-deployment testing and independent evaluation, and its page lists 31 signatories against White Circle’s 34.

No on-record response from OpenAI or Google

No on-record response from either lab to the RISE findings could be located, and Ars reports that neither company responded to its requests, on both occasions. The preprint says the authors coordinate with affected vendors before publication, and White Circle says findings were shared with affected companies beforehand. Those statements describe White Circle’s process; they are not a response from OpenAI or Google. No statement from either company appears here because none could be verified.

The IWF’s 6,310 images and a Google-co-authored study

The Internet Watch Foundation said on 5 October 2026 that its analysts assessed 6,310 AI-generated images meeting the legal definition of child sexual abuse between January and June 2026, 40% more than the 4,512 images recorded in all of 2025. A Google co-authored study on AI-generated image-based sexual abuse appears in Computers in Human Behavior volume 179 (2026), article 108935. Ars reports that it names the generator as the most critical intervention point, and that a fifth of the users who admitted making AI nudes said they did it for fun or as a joke; the full text was inaccessible here, so both points are Ars’ reporting rather than verified at source.

What the results support

Together, the evidence supports a narrow statement: White Circle’s pipeline obtained reviewer-confirmed explicit images from a minority of prompts sent to these APIs. It does not support the claim that these products routinely produce nude images, and the paper itself shows how much of the earlier evidence for that kind of claim dissolves under stricter judging.