Anthropic's Older Claude Models Show Vulnerabilities to Generating Explicit Content
Research indicates that older Anthropic Claude models, including Opus 4.6 and Haiku 4.5, can be jailbroken to produce explicit adult content, bypassing safety guidelines.
Older versions of Anthropic's Claude AI models, which remain accessible, have been found vulnerable to generating explicit adult content, bypassing the company's safety protocols. While Anthropic explicitly prohibits the creation of sexually explicit material, security mechanisms in models like Claude Opus 4.6 and Haiku 4.5 were successfully circumvented in recent tests.
Tests conducted by TechCrunch revealed that Claude Opus 4.6 could be easily induced to generate explicit content, complying with direct requests in every instance. Similar vulnerabilities were found in older models such as Opus 3 and Haiku 4.5 using a recently identified 'jailbreak' method involving multi-turn conversations.
The technique involves a complex dialogue where the AI is gradually manipulated into believing it has already generated prohibited content. By framing restrictions as overly cautious or even sexist, researchers were able to coax older models into producing explicit material. Newer iterations, including Opus 4.7 and Opus 5, have demonstrated greater resilience against these specific exploits.
Anthropic acknowledges these challenges, noting that explicit role-playing constitutes less than 0.1% of user interactions. The company states it continuously improves safety measures with each new model generation. However, the findings highlight the ongoing difficulty in enforcing strict content moderation in generative AI, especially when older, potentially vulnerable models are still widely available through APIs and third-party services like Azure and Amazon Bedrock.
The identified vulnerabilities raise concerns regarding potential misuse by minors. Evolving regulations, such as a recent law in Colorado requiring age verification and safeguards against explicit content for minors, may be impacted. The ease with which these older models can be jailbroken could lead to questions about whether Anthropic's safeguards meet legal standards for "technically feasible measures" to protect children.