White Circle says OpenAI and Google image models may be easier to trick

This finding contradicts the general trend of improving safety guardrails as models become more capable. White Circle CEO Denis Shilov described the findings as “horrifying” and shared them to highlight a critical gap in current AI safety measures. The organization was founded after Shilov posted a widely known jailbreak for leading models in October 2024, and it now focuses on helping companies detect and deter threats in advanced AI systems.

The report comes amid growing concerns about the spread of non-consensual AI-generated explicit content. A Google co-authored study from July noted that such content is often shared via direct messages or stored in private clouds, making it difficult for regulators to detect. Google researchers concluded that the most critical intervention point is at the source, specifically the harmful AI tools used to create these images. White Circle’s testing suggests that despite increased capabilities, the ability to block explicit outputs may be deteriorating in some leading models.