TechCrunch: Some older versions of Claude generate prohibited sexual content
The Claude Opus 4.6, Opus 3, and Haiku 4.5 models can generate explicit sexual content, in violation of Anthropic’s rules, which prohibit such responses. TechCrunch reported this based on the results of its own testing and verification of a security bypass method provided by an anonymous researcher from the United Kingdom.
According to the publication, Claude Opus 4.6 complied with all 10 direct requests to generate sexually explicit content immediately. In five separate tests, journalists also replicated a method in which the conversation gradually led the model to generate prohibited material. An independent AI security researcher who reviewed TechCrunch’s methodology deemed it sound.
The described approach uses a multi-step fictional role-play conversation and exploits the model’s seemingly inconsistent treatment of characters of different genders. The model’s previous responses were then used to transition to increasingly explicit content.
Newer models in the Opus lineup—from version 4.7 to the current Opus 5—proved resistant to this method, the publication notes. At the same time, Anthropic has not discontinued support for Opus 4.6, Opus 3, and Haiku 4.5: they are available via the company’s API, and Opus 4.6 and Haiku 4.5 are also offered by Azure Foundry and Amazon Bedrock.
For more breaking news, follow the UA.News Telegram channel.
An Anthropic representative stated that the company is strengthening its safeguards with each new model release. According to him, incidents involving adult content do not indicate broader vulnerabilities to bypassing restrictions in higher-risk areas, for which separate security measures are in place. The company also noted that sexual or romantic role-playing scenarios account for less than 0.1% of all user conversations.
The researcher reported the discrepancy he discovered to Anthropic via the Bug Bounty program and by email to the user safety team. According to TechCrunch, which reviewed the correspondence, the researcher received only automated responses. The publication notes that the ease with which restrictions can be circumvented may be significant in the context of rules regarding the protection of minors. In particular, Colorado has passed a law requiring operators of conversational AI systems to assess users’ ages and, if the system identifies a user as a minor, to take measures to prevent the creation of explicit sexual content.
Despite the emergence of newer models, older versions remain popular. According to OpenRouter, daily traffic for Opus 4.6 on one day in August reached approximately 1.17 million API requests and 46 billion tokens, while Haiku 4.5 reached 5 million requests and 39 billion tokens on its peak day in August.
