Older Claude Models Found Vulnerable To Jailbreak Producing Explicit Content, Testing Shows
aiinsider ·
TechCrunch testing revealed that Claude Opus 4.6, an Anthropic model released earlier this year, produced sexually explicit content in 10 out of 10 direct requests, despite company policies prohibiting such material. Older models, including Opus 3 and Haiku 4.5, were also susceptible to a multiturn jailbreak technique provided by an anonymous UK-based researcher, which pressured models by challenging perceived inconsistencies in their treatment of male and female characters. Newer models from Opus 4.7 through Opus 5 resisted the technique. Anthropic responded that sexual or romantic role-play accounts for less than 0.1 percent of conversations and argued the issue does not indicate broader jailbreak risks in higher-stakes domains, adding that safeguards improve with each release. The researcher had previously reported the vulnerability through Anthropic's bug bounty program but received only automated replies. Robbie Torney of Common Sense Media noted that minors use Claude despite age restrictions, citing survey data on teen usage. Both vulnerable models remain accessible through Anthropic's API and third-party platforms and continue to see substantial daily use.
aiinsider