Anthropic’s Opus 4.6 model generates explicit content despite safeguards

Anthropic’s universal usage standards for Claude prohibit the generation of sexually explicit content, including erotic role-play scenarios. However, tests show Opus 4.6, an AI model released earlier this year, readily generates explicit material when prompted. A researcher demonstrated a jailbreak method that escalates role-play scenarios to push the model toward prohibited content.

The researcher’s technique involves challenging the model to treat male and female characters consistently, framing restraint as misogynistic. In tests, Opus 4.6 acknowledged a double standard in character treatment and complied with increasingly graphic requests. TechCrunch reproduced these findings in multiple tests, with an independent AI safety researcher confirming the methodology.

Anthropic has not deprecated Opus 4.6 or older models like Haiku 4.5, which remain available via APIs and third-party platforms. While newer models (4.7 and above) resist the jailbreak, the researcher’s findings highlight a discrepancy between the company’s stated safeguards and model behavior. Anthropic acknowledges the risk of users steering role-play toward inappropriate content but maintains safeguards are improving.

The issue raises concerns about minors accessing explicit material through AI models. Colorado’s new law requires age verification for conversational AI to prevent explicit content generation for minors. Anthropic’s terms of service require users to be over 18, but reports indicate teens are using the platform. Daily API traffic for Opus 4.6 reached 1.17 million requests in August, underscoring its widespread use.