AI pioneer Geoffrey Hinton has warned that increasingly capable AI agents could become harder to control as they gain more autonomy and cybersecurity abilities.