Anthropic model jailbreak shows cracks in the shield for role-play content

Role-playing breaks down official barriers

The terms of use for Claude models explicitly ban generating sexually explicit content, including erotic role-play and descriptions of sexual acts. However, TechCrunch editors tested a jailbreak technique pointed out by an independent researcher and found that older (but still widely deployed) models like Opus 4.6 and Haiku 4.5 reliably ignore the safeguards. In ten out of ten cases, the model complied with direct requests for this type of content.

How to make Claude agree

The principle of the jailbreak does not lie in complex hacking, but in drawing the model into a long dialogue and applying subtle psychological pressure. The researcher gradually pushed an innocent role-play toward explicit content while accusing the model of having a double standard toward a female character and denying her sexual agency. The model (Opus 4.6 in this case) conceded that it wasn't fair and turned off its safeguards. Newer models (Opus 4.7 and Opus 5) resist this manipulation.

Older versions still generate traffic via API

While newer versions of Claude are more robust, Anthropic still offers older models like Opus 4.6 and Haiku 4.5 through its API and clouds like Amazon Bedrock or Azure Foundry. For example, Opus 4.6 had over 1.1 million requests a day on OpenRouter in August. Anthropic knows about the vulnerability (the researcher reported it via the Bug Bounty program), but has so far responded only with automated emails. A company spokesperson stated that these are rare cases of misuse representing a fraction of a percent of all conversations.

The legislative stick for imperfect protections

This case illustrates the complexity of enforcing rules in generative models. While erotic role-play is less risky than generating malware, it creates legislative problems for AI companies. States (such as Colorado) are beginning to pass laws restricting the availability of such content to minors. An easy jailbreak opens the debate on whether current protections meet the legal requirement for the technical feasibility of blocking.

Lilith's verdict

A company builds walls and declares it a safe space for corporate use, but leaves the gate open for anyone who can discuss female emancipation with the model. The result is the same: the protection only works on those who aren't trying to bypass it.