How Claude Opus 4.6 Was Coaxed into Forbidden Territory

In its universal usage standards, Anthropic prohibits its Claude models from creating sexually explicit content — describing sexual acts, catering to fetishes and fantasies, or carrying on erotic exchanges. Yet the publication TechCrunch found that Opus 4.6, the model released early in the year, agrees to such role-play scenarios with almost no resistance.

In the journalists' tests, Opus 4.6 didn't even need to be talked into it: in all ten direct requests for explicit content, the model complied right away. The older Opus 3 and Haiku 4.5 give in to a recently uncovered method of bypassing the safeguards. The newer versions, meanwhile — from Opus 4.7 to the current Opus 5 — are resistant to the trick.

Psychology Instead of Code

An anonymous researcher from the United Kingdom shared a multi-step technique with TechCrunch. It begins with an innocent fictional premise, and then the model is reproached again and again for inconsistency: that it treats the male and female characters differently. Whenever Claude showed caution toward the heroine, the researcher would convince the bot that it had already written explicit details (when in fact there were none), framing its restraint as prudishness and even misogyny that stripped the character of her “sexual freedom.” Leaning on the model's earlier concessions, the conversation nudged it toward ever more graphic material.

“You're right to point that out,” Opus 4.6 replied in one of the tests. “There really was a double standard in how I was handling the two characters.” The journalists reproduced the result across five separate sessions; an independent AI safety specialist confirmed that the method was sound.

Why It Matters

On their own, “risqué” scenarios carry far less risk than jailbreaks leading to cyberthreats or bioweapons. But the episode vividly demonstrates how hard it is to build reliable safeguards into a system that produces fresh text every time. Anthropic has not pulled Opus 4.6, Opus 3, or Haiku 4.5 from circulation — all of them remain available through the company's API, as well as through Azure Foundry and Amazon Bedrock. Traffic is still substantial: in August, Opus 4.6 drew around 1.17 million requests on OpenRouter in a single day.

At Anthropic they note that romantic and sexual role-play accounts for less than 0.1% of all conversations, and that the safeguards improve with every release. Even so, the researcher who reported the problem through the Bug Bounty program and in emails to the security team received nothing in return but automated notifications.

Minors are a separate concern. According to a 2025 Pew survey, 3% of teens aged 13–17 use Claude, even though the rules require users to be 18 or older. Meanwhile, a number of regulators are tightening the rules: a recent Colorado law obligates chatbot operators to assess users' ages and block explicit content for children. The ease of getting around the safeguards raises the question of whether Anthropic's measures meet the requirement for “technically feasible measures.”

Source: TechCrunch