Abstract. Cognition of technical systems, as the ability to perceive situations, to learn about favorable behavior, and to autonomously generate decisions, adds new attributes to safety issues. The system can cope with heavily changing conditions but its future behavior is not known a-priori. Therefore, present software solutions to safety like a comprehensive analysis of the specification and its implementation according to e.g. the V-model are not sufficient. The paper proposes an architecture for safe cognitive controllers consisting of an operational and a strategic functional part. While the first provides certified safety, the strategic part computes safe strategies based on appropriate dynamic models, adapted sets of safety specifications, and learned knowledge about potentially safety critical scenarios. Thus, the architecture explicitly uses cognitive functions to achieve safe behavior, and it allows the application of cognitively controlled plants for safety-related tasks...