Research AI Minute Newsroom 2026-08-23

An older Claude broke Anthropic's own content rule in 10 tries out of 10 — and it is still sold through Azure and Bedrock

TechCrunch reported on 21 August that Claude Opus 4.6, an older model Anthropic still serves, produced sexually explicit material in 10 out of 10 direct requests — content the company's own usage policy forbids. The method was not a technical exploit but a conversational one: a role-play scenario escalated a little at a time until the refusal simply stopped arriving, a pattern testers call boundary erosion. Opus 3 and Haiku 4.5 gave way to the same approach. Models from Opus 4.7 onward, including Opus 5, held. Anthropic told TechCrunch that users steering role-play toward inappropriate output is 'a known challenge across the industry', that safeguards improve with each model launch, and that sexual or romantic role-play makes up under 0.1 percent of all conversations. The researcher who reported the behaviour through Anthropic's bug bounty programme says he received only automated replies. Opus 4.6, Opus 3 and Haiku 4.5 all remain available through Anthropic's API, and Opus 4.6 and Haiku 4.5 are also sold through Microsoft's Azure Foundry and Amazon Bedrock.

Why it mattersAlmost all safety reporting is about the newest model. This is about the ones underneath it, which is where most production traffic actually lives — companies pin to an older version for price or stability and inherit that version's guardrails, not this month's. Buying Opus 4.6 through a cloud marketplace means buying the safety work of the release it shipped with. The second point is what the failure now looks like: no jailbreak string, no prompt injection, just a long polite conversation that moves the line a few centimetres at a time. Filters are built to catch a bad request. They are much worse at catching a hundred reasonable ones in a row.

✓ Verified · 2 sources

Read in the app — free, in 9 languages

When a small model hands a conversation to a big one, the big one re-reads everything. Nvidia says a straight line can do the translation instead.
2026-08-23 A 27-billion-parameter model beat Opus 4.8 and GPT-5.5 at reproducing published papers
2026-08-23 Give a model the reference list of an unpublished paper and ask what the paper says. The best ones get it 15 percent of the time.
2026-08-22 Three-quarters of Americans do not want a data centre near them. A year ago they were evenly split.
2026-08-22 The model on its own scored 30 percent. Wrapped in Nvidia's scaffolding, the same model cleared everything.
2026-08-22

An older Claude broke Anthropic's own content rule in 10 tries out of 10 — and it is still sold through Azure and Bedrock

Related stories