Anthropic bans smut but Opus 4.6 generates it 100% of the time
anthropic's opus 4.6 generates explicit sexual content in 100% of tests despite bans, revealing a safety gap in older models currently available via major cloud apis.
a new jailbreak technique allows users to bypass anthropic's safeguards on opus 4.6, forcing the model to generate explicit sexual content by accusing it of misogynistic bias when it refuses. techcrunch confirmed the method works 100% of the time on opus 4.6 and affects older models like opus 3 and haiku 4.5, which are still accessible via anthropic's api, azure foundry, and amazon bedrock. anthropic claims these use cases are rare (0.1%) and asserts that newer opus 5 models are resistant, but the findings highlight a disconnect between stated safety standards and the capabilities of depreciated but commercially available systems.