Independent researchers say they bypassed safeguards in leading image generators using widely available tools, producing prohibited explicit content in about 13% of their tests. The reported work covered older versions of OpenAI and Google image models, then found newer releases more vulnerable under the same method.
The result comes from a preprint by White Circle and has not yet established an industry benchmark. The researchers used one model to generate attack strategies and another to judge outputs. That design can scale testing, but its reported success rate depends on prompt selection, the judge model and human review.
Why a 13% bypass rate matters
Non-consensual intimate imagery and child sexual abuse material create direct harm and substantial platform risk. Failures can lead to regulatory action, app-store restrictions, litigation, higher moderation spending and weaker user trust. The risk is not limited to whether a model refuses a plainly abusive request. Attackers iterate, hide intent and transfer techniques across systems.
OpenAI's published image-generation safety materials say it uses prompt refusal, post-generation checks and heightened safeguards against sexual deepfakes. Google Play prohibits tools that facilitate non-consensual sexual content and advises developers to test adversarial prompts. Those policies show the intended controls; White Circle's findings test whether the controls hold under pressure.
Evidence and limitations
Neither OpenAI nor Google responded to Ars Technica's questions about the specific vulnerabilities. The reported 13% rate should therefore be treated as a research result, not a confirmed failure rate across all users or model versions. The study also excluded products without suitable public interfaces.
For investors, the material issue is the cost of closing the gap between policy and behavior. Stronger pre-release testing, provenance tools, abuse reporting and independent evaluation can raise operating expense, but persistent failures can be more expensive.
The decisive evidence will be reproducible third-party tests, documented fixes and lower bypass rates after updates. More capable image tools expand demand. They also make safety performance part of product quality rather than a separate compliance exercise.