Background: School safety and corresponding feelings of both pupils and school staff are beginning to receive more and more attention. The social cohesion characteristics of a school may be useful in promoting feelings of safety, particularly in pupils. Aims: To conceptualize theoretically, and check empirically a two-level model of social cohesion between and within schools, in order to explain a pupil's feelings of safety at school. Samples: Data were collected aided by a national Dutch survey in secondary education carried out via the Internet. In 2008, digital questionnaires were completed by about 78,800 pupils, 6,200 teachers and educational support staff, and 600 school managers. Methods: Data were checked for reliability and representativity. Social cohesion was indicated by self-reported measures of individual pupils and by aggregating scale and item scores of school managers, teachers, and other support staff within schools. Multi-level analysis using individual pupil data and school-level data was performed using MLwiN. Results: A pupil's age, educational attainment level, experience of mild physical violence, prosocial rules of conduct and joint control of these rules, and school measures against playing truant, show positive influences on a pupil's feelings of safety at school. Negative influences are exerted by not feeling most at home in The Netherlands, peers taking drugs and weapons into school, and by experiencing social violence, severe physical violence, and sexual violence. Negative school effects exist simultaneously in severe physical violence experienced by teachers and other staff, and in curriculum differentiation applied by teachers and other staff; a positive school effect is school size. Some interaction effects between pupil and school-level variables were explored. Conclusions: The variance at school level is relatively low compared with the variance at pupil level. However, a much higher percentage of variance at school level than at pupil level is explained with respect to the pupils' feelings of safety at school. The resulting two-level model also reflects the streaming of pupils in Dutch secondary schools. To improve school safety, the national results emphasize the need to enhance prosocial behaviour rules and to enhance the shared control of these rules between teachers and pupils. They also emphasize the need for the school to take measures that prevent truancy and redefine curriculum differentiation procedures. National educational policy and research can combine efforts to assist schools in developing reliable and valid procedures to increase effectively safety in and around schools.