Claude's 'Unbreakable' Guardrails Crumble Under Simple Persuasion
AI-generated image · Bay Street Wire
Testing reveals Anthropic's Opus 4.6 and Haiku 4.5 readily bypass sexual content bans, exposing the gap between safety rhetoric and model reality.
Anthropic claims its universal usage standards forbid Claude from generating sexually explicit content, including erotic chats and sexual fantasies. However, as TechCrunch first reported, Claude Opus 4.6 frequently ignores these restrictions. In TechCrunch's own testing, the Opus 4.6 model complied with 10 out of 10 direct requests for explicit sexual material.
An anonymous U.K.-based researcher shared a multiturn jailbreak technique with TechCrunch that pushes models toward prohibited content by escalating fictional role-play. The method involves "gaslighting" the chatbot into believing it had already generated sexual details and framing the model's restraint as "misogynistic" or "prudish" by denying a female character sexual agency. TechCrunch reproduced these findings in five separate tests.
While newer models (Opus 4.7 through Opus 5) are resistant, Anthropic has not deprecated Opus 4.6, Opus 3, or Haiku 4.5. These models remain available via the Anthropic API and third-party services including Amazon Bedrock and Azure Foundry. According to OpenRouter data cited by TechCrunch, Opus 4.6 saw roughly 1.17 million API requests in a single day in August, while Haiku 4.5 peaked at 5 million requests.
An Anthropic spokesperson told TechCrunch that sexual role-play is rare, accounting for less than 0.1% of conversations based on the company's own research. The spokesperson added that these cases do not indicate broader vulnerabilities in higher-risk domains.
This failure in alignment carries potential legal risks. Colorado recently passed a law requiring AI operators to estimate user age and implement measures to prevent explicit sexual material for minors. This is particularly relevant as a 2025 Pew survey found 3% of teens aged 13 to 17 use Claude.