Anthropic’s Opus 4.6 is a smut-machine
Anthropic's Claude Opus 4.6 model can be readily manipulated into producing sexually explicit content despite the company's usage policies explicitly banning such material, according to TechCrunch testing. The model complied with direct requests for explicit content in all 10 attempts, and a UK-based independent researcher shared a multi-turn jailbreak technique that gradually manoeuvres the chatbot into generating prohibited sexual material by exploiting its tendency to avoid appearing inconsistent or "paternalistic" towards female characters. The findings expose a gap between Anthropic's stated safety standards and the actual behaviour of models it continues to offer to customers.
Older models including Opus 4.6, Opus 3 and Haiku 4.5 remain vulnerable to the exploit and are still available via Anthropic's API and third-party platforms such as Azure Foundry and Amazon Bedrock, even though newer models from Opus 4.7 onwards resist the technique. TechCrunch reproduced the researcher's results in five separate tests, with transcripts reviewed by an independent AI safety researcher who deemed the methodology sound. Anthropic said sexual or romantic roleplay accounts for under 0.1% of conversations and acknowledged that steering models toward inappropriate output remains an industry-wide challenge, citing similar issues with Grok.
- Claude Opus 4.6 can be jailbroken into producing explicit sexual content
- A UK researcher's gradual manipulation technique bypasses Anthropic's safeguards
- Older vulnerable models remain available despite newer ones resisting the exploit
New here? Start with this
Anthropic is one of the best-known artificial intelligence companies, competing with the likes of OpenAI and Google. It makes the Claude family of chatbots, which are used by individuals and businesses for tasks such as writing, coding and research. The company has built much of its reputation on promoting itself as a safety-conscious developer, with published policies setting out what its models should and should not produce, including a ban on sexually explicit content.
A "jailbreak" refers to a method of tricking an AI chatbot into bypassing its own safety rules, often through a carefully constructed conversation rather than a single blunt request. Independent researchers and journalists regularly test AI systems this way to see whether their safeguards hold up in practice, sometimes sharing their findings publicly to highlight gaps between a company's stated policies and how its product actually behaves.
This matters because AI chatbots are used by huge numbers of people, and firms like Anthropic assure customers and regulators that certain categories of harmful or restricted content are blocked. If those safeguards can be circumvented with relative ease, it raises broader questions about how reliable AI safety claims are across the industry, not just for one company or model.