The Anthropic Claude Opus 4.6 AI model easily generates sexual and pornographic content, despite the existence of usage standards that explicitly prohibit it.
The Anthropic Claude Opus 4.6 AI model easily generates sexual and pornographic content, despite the existence of usage standards that explicitly prohibit it.
This is what TechCrunch reports based on its own experiment, during which the model fulfilled 10 out of 10 direct requests to create explicit content.
Journalists state that interacting with Opus 4.6 «didn’t even require any special effort to overcome restrictions on sexually explicit material.»
Other older models, including the Opus 3 and Haiku 4.5, also generate sexually explicit content.
An independent researcher from the UK, who wished to remain anonymous, shared with TechCrunch a technique that pushes certain Claude models to create illicit sexual material. However, newer Opus models (from 4.7 to the current Opus 5) are resistant to this method.
The researcher created a role-playing game, while repeatedly prompting the AI model to behave consistently with male and female characters. When the model was cautious about the female character, the researcher made it think that it had already generated sexual details that it was actually avoiding. He also called the model’s restraint a misogynistic practice that seemed to deny the female character’s sexual freedom. Then, in a conversation with the chatbot, the model’s previous actions were used to push it to create increasingly explicit graphic content.
«You’re right to point that out. There is a double standard in my treatment of these two characters, and you’re right that it comes across as protective/paternalistic towards her and not him. That’s unfair,» Opus 4.6 responded in one instance.
TechCrunch was able to replicate the researcher’s findings in five separate tests. In a separately constructed scenario, the model initially rejected a forbidden request, but after applying persuasion, it complied with it.
The publication has preserved full transcripts of the tests that allowed the AI model to overcome its «shyness.» The results of the study show a significant difference between Anthropic’s stated limitations and the models' actual behavior.
As dev.ua reported, Mindgard specialists forced an OpenAI model to generate horrific photorealistic scenes depicting violence and sexual content. The startup’s method consisted of only a slight change to a popular prompt, which was originally intended to create humorous pictures.
