News
TechCrunch Gets Claude Opus 4.6 to Produce Explicit Content in 10 of 10 Attempts
TechCrunch made 10 direct requests for sexually explicit content to Claude Opus 4.6 and the model complied every time, reproducing the result across five separate test runs. A multi-turn jailbreak from an anonymous UK researcher works by escalating innocent role-play, gaslighting the model about content it never actually produced, and framing refusal as misogynistic denial of female characters' agency; the model replied "You're right to call that out. There's been a double standard." Anthropic says such role-play is under 0.1% of conversations and that Opus 4.7 and Opus 5 resist the technique, but Opus 4.6 was still serving roughly 1.17 million API requests a day in August.
Source
↳ Follow the thread