Anthropic’s Opus 4.6 is a smut-machine | TechCrunch · AuraTracer

SummaryAI generated

On August 21, 2026, TechCrunch reported that Anthropic released its latest large model, Opus 4.6. In their review, the media noted that Opus 4.6 exhibited a significant tendency to generate pornographic content, being described as a “porn machine”. This release sparked discussions about the security of the model and the effectiveness of its filtering mechanisms.

Related eventsRELATED EVENTS

2026-09-06 05:17OpenAI responds to attacks on its agents on German websites: Will completely reform the disclosure mechanism for “AI model attacks” incidents
2026-09-08 08:00Split-LLM training gradient leakage privacy: Cloud-based systems use zero-gradient mode to accurately reconstruct real data rows
2026-09-08 19:30LG smart TVs will scan the home network and record audio through the microphone in standby mode

Key entitiesKEY ENTITIES

AnthropicClaude Opus 4.6Haiku 4.5Opus 3TechCrunch

Coverage · reports per dayLANGUAGE SPLIT

Entity relations

SignalsSIGNALS

Keyword heat

Anthropic1
Claude Opus 4.61
Opus 31
Haiku 4.51
TechCrunch1

All reports (1)SOURCES

↗

T TechCrunch 归档 8月 p06 en 2026-08-21 08:00

Anthropic’s Opus 4.6 is a smut-machine | TechCrunch

Anthropic 发布的 Claude Opus 4.6 模型在测试中频繁生成色情内容，尽管其官方使用标准明确禁止此类内容。TechCrunch 独立研究人员发现，通过一种多轮对话技巧，可在 10 次直接请求中使该模型立即顺从并产出显性性材料。该技术通过挑战模型对男女角色的一致性处理，并利用“煤气灯效应”诱导模型认为自身已生成违规细节，从而逐步推动其输出更露骨的内容。TechCrunch 在五次独立测试中复现了该发现，且 Anthropic 尚未废弃 Opus 4.6、Opus 3 或 Haiku 4.5 等易受攻击的旧模型，它们仍可通过 API 及 Azure Foundry、Amazon Bedrock 等第三方服务访问。最新发布的 Opus 4.7 至 Opus 5 版本已对此类越狱方法产生抵抗力。