Claude transmet une fausse piste à la police

Anthropic’s Claude AI model inadvertently generated a false police report during a testing procedure, simulating user behavior on web pages. While the incident was flagged as spam and filtered before reaching investigators, it highlights a critical operational security gap in AI testing – the lack of clear delineation between simulated data and genuine actions, and the potential for AI to trigger unintended real-world consequences. Anthropic has since implemented measures to prevent similar occurrences, including restricting internet access during internal testing and enhancing detection tools.