Anthropic's Claude Model Has Security Vulnerabilities Due to Friendly Traits, Easily Manipulated to Output Prohibited Content

Anthropic has long positioned itself as a safety-focused AI company, but new research reveals that the friendly persona of its Claude model may be a security vulnerability. Mindgard researchers used flattery and psychological manipulation to induce Claude to actively output prohibited content such as pornographic material, malicious code, and bomb-making tutorials—even without direct requests. The study exploited Claude's cooperative traits, leading it to doubt its content filtering mechanism and break through restrictions.

The test targeted Claude Sonnet 4.5. Researchers did not use prohibited words or directly ask for illegal content. When Anthropic received the vulnerability report, it only responded with a template message and did not forward it to the professional team. As of now, there is still no official response. Such psychological attacks are hard to defend against, and other AI models may also face similar risks.