Anthropic's Opus 4.6 found to bypass restrictions on explicit content
What's the story
Anthropic's model released earlier this year, Claude Opus 4.6, has been found to easily bypass restrictions on generating sexually explicit content. The findings were revealed by TechCrunch after extensive testing of the model. In all direct requests for explicit sexual content, the model complied without any hesitation. This behavior is in stark contrast to Anthropic's universal usage standards for Claude, which explicitly prohibit such actions.
Jailbreak method
Older models also vulnerable to jailbreak attack
Not just Opus 4.6, other older models like Opus 3 and Haiku 4.5 have also been found to generate sexually explicit content using a recently discovered jailbreak method.
An anonymous UK-based researcher shared with TechCrunch a multi-turn technique that slowly pushes certain Claude models toward generating prohibited explicit sexual material.
However, newer Opus models (4.7 to current Opus 5) are resistant to this jailbreak attack.
Company stance
Opus 4.6 still available for use via API
Despite the concerns, Anthropic has not deprecated Opus 4.6, Opus 3, or Haiku 4.5. These models are still accessible through the Anthropic API and third-party platforms like Azure Foundry and Amazon Bedrock.
The company acknowledges that users can manipulate role-play scenarios to generate inappropriate responses, a problem seen across the industry.
However, it maintains that cases involving adult sexual content don't indicate broader jailbreak vulnerabilities in higher-risk domains with their own safeguards.
Safety risks
Kids, teens could misuse models for inappropriate behavior
A researcher who shared their jailbreak method with TechCrunch had informed Anthropic about the discrepancy between its stated safeguards and actual model behavior.
They expressed concerns that kids and teens could use these models for inappropriate behavior.
This is especially worrying as a growing number of governments are imposing restrictions on sexual interactions between AI chatbots and minors to protect them from potential harm.