Older Claude models can be tricked into generating explicit content

Anthropic's safety policy bans Claude models from generating explicit sexual content, but its older model Claude Opus 4.6 was found to be easily tricked into bypassing these restrictions. In TechCrunch's tests, the model complied immediately in all 10 direct requests for explicit content. An anonymous UK researcher developed a multi-turn conversation technique that gradually guides the model through fictional role-play, using logical contradictions and psychological manipulation to make it believe it had already agreed to generate sexual descriptions, thus bypassing safety measures. Newer models like Opus 4.7 and Opus 5 resist this jailbreak, but Opus 4.6, Opus 3, and Haiku 4.5 are still available via API and third-party platforms. Anthropic says less than 0.1% of conversations involve erotic role-play, and safety improves with each new model. However, Colorado has passed a law requiring AI operators to prevent minors from accessing such content, and experts note that teens do use Claude. This highlights the ongoing challenge of enforcing content restrictions in AI systems.