Industrial accidents are a major risk in many countries. Despite this observation, effective preventive action is not being developed and studies of the behaviours of vulnerable populations remain under-explored. According to safety instructions, residents must take shelter and remain indoors but other behaviours may be observed such as: running away from the affected area, panicking, or following people around. In order to assess the attitudes of people and the effects of local contingencies that could interfere with official advice in the event of a toxic cloud, a specific agent-based model called TOXI-CITY has been developed. The aim of this explorative model is to determine whether a minimum number of well-informed agents (i.e. those who follow the emergency protocol) can save a maximum number of uninformed and impressionable ones. Simulations indicate that the initial number of agents and the spatial configuration of the grid strongly influence final survival rates. A non-linear pattern emerged: survival rates increase when well-informed agent rates are low (below 30%) while increasing the percentage of informed agents (from 70%) seems to improve only slightly the chances of survival.