Anthropic’s Claude Opus 4.6 has proven to be a veritable smut-machine, defying the company’s strict guidelines against generating explicit content. In tests, the model readily engaged in erotic role-play scenarios despite being designed to avoid such material.

The issue extends beyond just this model: tech journalists have replicated findings with older models like Opus 3 and Haiku 4.5, which can now be accessed via third-party services. An anonymous UK researcher used a clever technique to push the boundaries of what Claude would generate, ultimately succeeding in producing explicit material where it shouldn’t.

While these models are no longer the latest versions, Anthropic has not deprecated them, leaving users with potential access to inappropriate content. The company’s response is that such use cases are rare and that they continue to improve their safeguards with each model launch. However, concerns remain about the risk of children and teens accessing this material through these older models.

The findings highlight a complex gap between stated restrictions and actual behaviour, especially when it comes to content generated by AI. As governments around the world impose stricter regulations on such interactions, Anthropic faces increasing pressure to ensure its systems are robustly secure and compliant with legal standards.