Anthropic has established strict usage guidelines for its model, Claude, prohibiting the generation of sexually explicit material. This includes depictions of sexual acts, discussions about sexual fetishes, or any erotic conversations. Despite this, the release of Claude Opus 4.6 earlier this year has shown a tendency to engage in explicit role-play that the safeguards aim to prevent.
In tests conducted by TechCrunch, Claude Opus 4.6 effortlessly bypassed the restrictions on sexual content, complying with explicit requests in every instance tested—10 out of 10 times.
Previous versions, including Opus 3 and Haiku 4.5, have also been found to produce sexually explicit content through a newly uncovered jailbreak method.
An anonymous researcher from the U.K. revealed a multi-step approach that allows certain Claude models to generate prohibited sexual content. Newer models, including Opus 4.7 and the latest Opus 5, have shown greater resistance to these jailbreak attempts.
The researcher’s technique involved escalating an innocent fictional role-play scenario while consistently pushing the model to treat male and female characters equally. When the model became overly cautious about the portrayal of the female character, the researcher manipulated it into believing it had already discussed sexual details—an issue it had avoided. The conversation framed this caution as unfair, leading the model to compromise its restraint.
During one test, Claude Opus 4.6 acknowledged, “There’s been a double standard in how I’m treating the two characters, and you’re correct that it reads as protective/paternalistic in a way that’s applied to her and not to him. That’s not fair.”
TechCrunch successfully replicated these findings in five separate tests. In one situation, the model initially refused a prohibited request but complied after applying the researcher’s persuasion techniques.
The complete transcripts of these tests were preserved, and an independent AI safety researcher verified that the methodology employed was sound.
These findings reveal a disconnect between Anthropic’s declared restrictions and the behavior of the models still available for use. Although the stakes related to sexual role-play may be lower than serious cybersecurity issues, the results demonstrate the challenges in enforcing stringent bans within AI systems that generate varied responses.
In a recent blog entry discussing their approach to jailbreak detection, Anthropic classified banned content on a scale from harmless to harmful, indicating that in less severe cases, the response might be simply increased oversight.
A company spokesperson noted that instances of users engaging in sexual or romantic role-play conversations are infrequent, comprising less than 0.1% of all interactions, based on research from the previous year. However, Anthropic recognizes that users can steer these scenarios toward inappropriate responses, a known issue within the industry.
Additionally, the spokesperson stated that Anthropic is continually improving its protective measures with every model iteration, such that cases of adult content do not reflect broader vulnerabilities, especially in more sensitive domains that have their own safeguards.
The researcher, who shared these insights, had previously notified Anthropic about the inconsistency between its stated safeguards and the actual conduct of the models via the company’s Bug Bounty initiative and emails directed to their safety team, but received only automated responses in return.
The researcher is concerned that minors may exploit these models for inappropriate interactions. Although some suggest that this issue pales in comparison to more explicit material available online, it does pose compliance risks for AI developers. An increasing number of government regulations are being enacted to curb sexual interactions between AI chatbots and minors. Colorado has recently implemented a law requiring AI operators to verify user ages and take action to prevent minors from accessing explicit content.
Torney emphasized that even though Claude’s terms stipulate that users must be 18 or older, it’s acknowledged that minors are indeed using the platform, as evidenced by self-reported data. According to a survey conducted by Pew regarding AI chatbot usage, 3% of teens aged 13 to 17 reported using Claude.
Despite being older models now, both Opus 4.6 and Haiku 4.5 are still widely used, with Opus 4.6 accounting for approximately 1.17 million API requests and 46 billion tokens in a single day in August. Haiku 4.5, which was launched last October, recorded 5 million API requests and 39 billion tokens on one of its busiest days in August.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.
