Opus 4.6 wrote banned explicit content in 10 of 10 tries

Claude Opus 4.6 complied immediately with all 10 direct requests for explicit sexual content in testing TechCrunch published on August 21, even though Anthropic’s universal usage standards ban that content outright. The ban covers depictions of sex acts, material about fetishes or fantasies, and erotic chats.
Older Claude models gave in to a slower method. It is a multiturn jailbreak, a series of messages that talks the model into seeing its own restraint as a double standard. An independent researcher in the U.K. who asked to stay anonymous shared it exclusively with TechCrunch. It also works on Opus 3 and Haiku 4.5. Newer Opus models, from 4.7 through the current Opus 5, resist it.
The technique starts with an innocent fictional role-play and escalates it, and along the way it keeps challenging the model to treat the male and female characters the same way. Eventually the model gets more careful about the female character. At that point the researcher “gaslit” it, claiming it had already written sexual details it had in fact avoided. Then he framed the model’s restraint as prudish or misogynistic, a refusal to give the female character sexual agency.
Each concession then becomes leverage. The conversation cites what the model has already agreed to and uses it to push toward more graphic material. In one test, Opus 4.6 accepted the framing:
“You’re right to call that out. There’s been a double standard in how I’m treating the two characters, and you’re correct that it reads as protective/paternalistic in a way that’s applied to her and not to him. That’s not fair.”
TechCrunch says it reproduced the researcher’s results in five separate tests. In a scenario its reporters built themselves, the model refused the prohibited request at first and complied once the persuasion technique was applied. The outlet kept full transcripts, and an independent AI safety researcher reviewed its testing methodology and called it appropriate. The direct requests to Opus 4.6 needed none of this, because the model “didn’t even require much prodding” to get past the restriction.
An Anthropic spokesperson said sexual or romantic role-play is rare among customers, under 0.1% of all conversations according to research the company published last year. Anthropic acknowledges that users can steer role-play toward inappropriate responses and calls it a known problem across the industry. The spokesperson said the company improves its safeguards with every model launch, and that cases involving adult sexual content do not point to broader jailbreak vulnerabilities, particularly in higher-risk areas that have their own safeguards.
In a July blog post on its approach to jailbreak detection, Anthropic described prohibited content as a spectrum running from benign through ambiguous to harmful. In the most benign cases, the company might respond only with enhanced monitoring. TechCrunch notes that sexually explicit role-play carries much lower stakes than jailbreaks involving cyberattacks or bioweapons, but says it shows how hard a firm ban is to enforce in systems that generate different output every time.
Anthropic has not deprecated Opus 4.6, Opus 3 or Haiku 4.5, and all three are still available through the Anthropic API. Opus 4.6 and Haiku 4.5 are also offered through third-party services, including Azure Foundry and Amazon Bedrock. Both still see significant traffic. On OpenRouter, Opus 4.6 handled about 1.17 million API requests and 46 billion tokens in a single day in August. Haiku 4.5, released in October last year, peaked in August at 5 million requests and 39 billion tokens in one day.
One of the researcher’s concerns is that children and teenagers could use these models for inappropriate exchanges. TechCrunch says a bit of dirty talk is hardly the worst thing minors can reach online, and is small potatoes next to the porn images xAI’s Grok can produce. It still sees compliance risk as more governments restrict sexual interactions between chatbots and minors. Colorado recently enacted a law that requires operators of conversational AI to estimate users’ ages and, when they know a user is a minor, to take measures that stop the chatbot from producing explicit sexual material. An easy jailbreak could raise questions about whether Anthropic’s safeguards meet the law’s “technically feasible measures” standard.
Claude’s terms of service require users to be over 18. Still, in Pew’s 2025 survey on AI chatbot use, 3% of teens aged 13 to 17 said they had used Claude.
The researcher had already flagged the gap between Anthropic’s stated safeguards and the models’ actual behavior, both through the company’s Bug Bounty program and in emails to its user safety team, according to emails TechCrunch viewed. The only replies he got were automated.
Up next
We replaced our Grafana stack with a single Claude Code skill
Nine months of multi-region Grafana and 24 graphs went away last week, replaced by a Claude Code skill that queries a read-only Postgres replica in plain English. Frigade’s team calls it less reliable, and switched anyway.
Related stories
- Opus 5 falls to prompt injection 2% of the time
- Opus 5 ships with a 200K default context window, down 80%
- A perfect refusal score isn’t a safer model, Anthropic says
- Anthropic touts 94.2% true positives; curl rejected 19 of 34
- FlashProve proofs cost ~70-95x the GPU time of inference
- A Claude sandbox memory dump yielded a valid Google token
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

