Wehrspohn Risk Management

How do you assess cyber risks with a Monte Carlo simulation?

Case study

By Dr. Uwe Wehrspohn.

A cyber risk arises from attacks from outside and from operational problems. Only some of them actually cause damage: there are direct losses and, with a certain probability, secondary losses that are often much larger. We break a cyber risk down into tangible factors, assess it with a Monte Carlo simulation in Excel, combine many cyber risks into a risk cluster and transfer the result into enterprise risk management.

For many companies, data and working computer systems are an essential part of their assets. Cyber risks potentially threaten these assets. A cyber risk materialises over several stages: it starts with contact with a problematic task or situation. This can give rise to a threat, the threat can give rise to a crisis, and the crisis can lead to losses.

Below we show how to break a cyber risk down into its factors, assess these factors and put them back together in a model in Excel. We then assess a single cyber risk with a Monte Carlo simulation, combine many cyber risks into a risk cluster and transfer the result into enterprise-wide risk management.

How do you go about assessing a cyber risk?

The approach is not limited to cyber risks. It is a general structuring principle that takes you from abstract risk considerations to very tangible, concrete factors. It consists of five steps:

  1. Identify the factors: in the first step, the relevant risk factors are identified, that is, the influences on the risk. They are, so to speak, the ingredients of the recipe.
  2. Assign magnitudes: in the second step, the factors are given magnitudes. Because they are now concrete rather than abstract factors, they can in some cases even be measured in the real world. If the company’s data is not yet developed to the point where direct measurement is possible, individual factors can still be assessed with expert estimates. The big difference is that concrete factors are now being assessed, factors you can picture, rather than abstract factors made up of many other factors.
  3. Describe how they interact: in the third step, we define how the factors are related and how together they produce the size of the risk.
  4. Evaluate: in the fourth step, we see what comes out for the specific risk.
  5. Analyse: a fifth step is particularly important for cyber risks. You can run stress tests, examine which factors matter most for the cyber risk and where most of the risk comes from, and combine many cyber risks.

Combining them is necessary because there is a very large number of cyber risks. Enterprise risk management, that is, company-wide risk management, cannot take in thousands of individual cyber risks. They have to be condensed and combined into clusters so that it can handle them at all.

What are the building blocks of a cyber risk?

Figure 1 shows the five factors that determine the size of a cyber risk: the attack frequency, the vulnerability, the size of the direct losses, the probability of secondary losses and the size of the secondary losses.

Figure 1: The five factors of a cyber risk from the slide “Relevante Faktoren: Was beeinflusst die Größe von Cyber Risiken?” (relevant factors: what influences the size of cyber risks?). The word “attack” is deliberately in quotation marks (German slide).

The first factor is the attack frequency. “Attack” is deliberately in quotation marks here because it does not always have to be an attack from outside. An internal operational problem can also lead to a cyber risk, for example when running the data centre with updates and system reboots, or simply through computers that break down.

The second factor is the vulnerability. Not every attack leads to an actual loss. Sometimes things turn out well and there is no loss at all. The vulnerability describes how the company responds to the attacks, that is, how many of them actually cause damage. It is therefore the second important key figure.

The third factor is the direct losses. They occur immediately when an attack is successful. Finally, there are secondary losses. They do not always occur but are only triggered with a certain probability. That is why two factors belong to them, their probability and their size. Secondary losses are often connected with something becoming public. This is not necessarily so. If a problem occurs and can be solved entirely internally, so that nobody outside the company learns about it, no secondary losses arise either. If they do occur, however, they are often much larger than the direct losses because the lever being pulled there is much longer.

Which attacks and errors trigger cyber risks?

Some attacks arise from crime. This ranges from hacking, data theft and fraud to phishing, in which fake e-mails, websites and the like try to obtain sensitive information. There can be unauthorised access or the disclosure of personal data. The actors behind them are just as varied. They can be criminals who want to make money from it, a bored student who attacks a company, or state-level actors who deliberately try to paralyse companies. This applies in particular to critical infrastructure.

Much more harmless and much less conspicuous “attacks” result from operational problems. Here you trip over your own feet, so to speak, without any outside involvement. A great classic is updates and restarts. They can paralyse the system, change features, introduce bugs, leave open security gaps they were actually supposed to close, damage data or cause incompatibilities. In practice, operational problems are often even much more frequent than attacks from outside, at least than the attacks from outside that are detected.

What do direct and secondary losses consist of?

Direct losses occur immediately when there is a problem. There are several classic examples:

  • If the system is paralysed, production first comes to a halt because the infrastructure is no longer available. This can also last for a longer period.
  • There are costs for restoring the system and, where necessary, the data if it was affected as well.
  • It may be necessary to investigate what caused the problem and how large the security incident was.
  • Emergency measures have to be taken.
  • There has to be communication, internally and possibly also externally with the authorities.
  • Where necessary, hardware and software have to be replaced.

To know the concrete consequences of a cyber risk, you therefore essentially have to follow up with a cost calculation. It distinguishes which items arise and how large they can be. This part is very commercially oriented and close to the practical work of the company.

The same applies to the secondary losses. If something becomes public, which is the classic example, the company may lose revenue because customers or business partners leave. A loss of reputation can lie behind this. But it can also be that customers and business partners could not obtain a service they needed and therefore had to find another provider. Insurance premiums, legal and compliance costs and other items can be added to this. Here, too, it is not set in stone which items make up the secondary losses. You have to look in detail at which costs arise in which situation.

Which distributions suit cyber risks?

Once the factors for occurrence, frequency and impact have been identified, you need a language for them. This language consists of giving each factor a distribution and choosing the parameters of that distribution. A distribution describes which values an uncertain factor can take and how likely they are. Figure 2 shows distributions that are typically well suited to cyber risks, in two groups: occurrences and frequencies at the top, impacts at the bottom. The situation is on the left, the matching distributions on the right. The two groups tend to use different distributions, although there is some overlap.

Figure 2: Distributions for cyber risks. Occurrences and frequencies at the top, impacts at the bottom, each with the situation on the left and the matching distributions on the right (German slide).

More on this: Which distributions are typically used, what properties they have and how to set their parameters is described in our guide “When Do I Use Which Distribution?” by Dr. Uwe Wehrspohn and Prof. Dr. Dr. Dietmar Ernst.

Occurrences and frequencies

The simplest case is that you know exactly in advance how often an event will occur. This is the case, for example, when there is a fixed operating plan for the system under which the server is restarted once a month. This happens twelve times a year, and the frequency is a constant number. This is of course a very special case.

If an event either does not occur at all or occurs exactly once, it follows a Bernoulli distribution. The Bernoulli distribution is a simple coin toss, but with a coin that does not necessarily show heads with a probability of 50% but with any probability. We can therefore freely decide with what probability an event should occur and then toss this coin.

If an event can occur between zero and n times a year, a binomial distribution may be suitable. It generalises the Bernoulli distribution to tossing the coin n times: in a single draw, it tosses a Bernoulli-distributed coin n times and counts how often heads came up. In practice, bounded distributions in which you specify a minimum and a maximum are also often used, for example the PERT distribution. Its three parameters are the minimum, the most likely value, also called the mode, and the maximum. The PERT distribution can take any number between the minimum and the maximum. With suitable rounding to whole numbers, however, it can also be used to describe frequencies. In both cases there is an upper limit: the event cannot occur more than n times.

If you cannot say so precisely how often an event can occur, it is an open multiple event. The Poisson distribution is the classic choice for this. It is again a generalisation of the Bernoulli and binomial distributions, namely for very large n, that is, for an upper limit that keeps growing until it practically no longer binds. Here, too, there cannot be fewer than zero occurrences, but the distribution is open to the right.

An alternative that often lends itself in practice is the Expert distribution. It has many more degrees of freedom, and in particular it allows extremely long tails. In everyday life you then mostly see relatively small frequencies, but it is not ruled out that extremely large numbers occur in some cases. This is an interesting model especially for criminal activity. A company may not have been in the sights of any criminals for many years and suddenly be seen as a target, so that the number of attacks rises sharply. Beforehand, this was expected only with a small probability, but it cannot be ruled out.

Instead of the Expert distribution, the Poly distribution, the lognormal distribution or the Weibull distribution can also be considered. All these distributions can represent two regimes, so to speak: with high probability you see small numbers, with small probability very large ones. Their parameters can be set both with expert estimates and with data, if data is available.

Impacts

On the impact side there is a similar distinction between cases. Bounded impacts have traditionally been described very well with the PERT distribution, the triangular distribution or the uniform distribution. The uniform distribution spreads the probability evenly between a minimum and a maximum. The Expert distribution and the Poly distribution, two newer distributions, are also very well suited here because they have variants that can be bounded. For both you can enter a minimum and a maximum if needed, and then the impacts are actually bounded.

Impacts can also be open, though. If the maximum consequences of an event are hard to grasp, it makes sense to assume that the impacts are open at the top. Then you need special distributions such as the Expert, Poly, lognormal or Weibull distribution. These are very similar considerations to those for the open multiple event.

Total impact: the compound distribution

Finally there is one more complication. An event occurs with a random frequency, so we do not know how many times it will actually happen, and each occurrence has a random size. Two kinds of randomness therefore come together. The compound distribution brings these two sides together. It takes the random number and the random impact, draws once from the impact distribution for each individual occurrence and adds up these values. The result is the total impact, that is, the sum of the losses from all occurrences.

How is the model in Excel structured?

In short: Risk Kit is an add-in for Microsoft Excel that lets you build and analyse Monte Carlo simulations directly in the spreadsheet. What we show here applies in general, though: assessing cyber risks with a Monte Carlo simulation works on the same principle in any simulation tool.

We now transfer the factors into a model in Excel. The template for it is included in Risk Kit from version 7.17. In a Monte Carlo simulation, such a model is recalculated very many times in a row with new random numbers. Each simulation run stands for one possible year. Together, all runs produce the distribution of the losses, from which the results are calculated.

Figure 3 shows the model for a single cyber risk before the simulation has run. It consists of three blocks. In the upper block, “Eingaben” (inputs), rows 5 to 9 contain the five factors from Figure 1, one factor per row. Columns C to H contain the parameters, column I the distribution we assume. All amounts in the model are in thousands of euros, written as T€ in the table.

Figure 3: The model for a single cyber risk before the simulation. Inputs at the top, calculations in the middle, results at the bottom, which are only defined after the simulation (German interface).

The middle block, “Rechnungen” (calculations), in rows 13 to 18 records how the factors interact. This is the third step of the approach. It contains the number of attacks, which follows from the attack frequency, the frequency of successful attacks from attack frequency and vulnerability, and the frequency of secondary losses, which also takes their probability into account. Then come the size of the direct losses and the size of the secondary losses. For each of them, the frequency drawn before and the distribution for the size of the losses play a role. Together they add up to the total losses in row 18. The values in this block are random numbers from a single run of the model. In Figure 3, for example, 52 attacks lead to two successful attacks and two secondary losses, and the total losses amount to T€ 15,278.

The lower block, “Ergebnisse” (results), in rows 22 to 30 evaluates the total losses: the probability that such a cyber risk occurs at all, the expected loss, various values at risk from the median to the value at risk at 99.9%, and the maximum of the distribution. Before the simulation, it still says “Erst nach der Simulation definiert” (only defined after the simulation), because these statistics need the results of all simulation runs.

How do you assess the attack frequency?

Let us go through the inputs one by one. In the example, we assume that a criminal attack on our system is to be feared. So we do not know in advance how often we will be attacked, and we cannot state a maximum for the attack frequency either. We therefore describe the attack frequency in row 5 with an Expert distribution.

The Expert distribution is defined by quantiles. A quantile is a value that is not exceeded with a given probability. The minimum in column C is zero because we cannot be attacked fewer than zero times. The good case in column D is ten attacks a year. The number of attacks falls below it only with a probability of 0.5%, as stated in column H under “Ws” for Wahrscheinlichkeit (probability). The median in column E is 30 attacks a year. The median is the middle value: with a probability of 50% there are fewer attacks, with 50% more. The bad case in column F is 100 attacks. The number of attacks stays below it with a probability of 99.5%, so it exceeds it only with a probability of 0.5%. The 0.5% from column H therefore applies to both sides. Column G, “Maximum”, is left empty, so the distribution is open to the right.

Data or expert judgement?

To estimate these parameters, you can use data the company may have, or you can have experts estimate them. If you have historical observations of how often the company has been attacked, these statistics tend to be very well suited for the estimate, because they are not extreme but moderate statistics. The minimum of zero does not need to be estimated; it is a given. A realistically good case, a realistically bad case and the median, that is, the middle of the observed distribution, can be read from the data. How things continue to the right along the extreme path does not matter at all for these three statistics; the median is not shifted by it at all. The only point is that, with a very small probability, you can still end up above the range you have observed.

Alternatively, you can assess the data with an expert judgement. Then the expert has to say what they consider realistic. The bad case may be something that has not yet been observed in the company. You may want to be deliberately conservative because you may not have detected all attacks. Perhaps there were blind attacks that nobody even noticed had taken place. Then you are cautious and raise the quantile for the bad case. This is a sensible correction of existing data by an expert judgement, a mixture of data analysis and expert judgement. The maximum is still left open.

In a different situation, for example when installing updates, you know from the outset that a certain number will not be exceeded, perhaps because there are not that many time windows for updates in a year. The exact number may not be known, but it may lie between 4 and 8. Then, too, you can use the Expert distribution, only with the minimum and maximum filled in and with corresponding quantiles. This is a great advantage of the Expert distribution: it is so flexible that it can be used for bounded as well as unbounded events.

More on this: How to define the Expert distribution through bounds, which may also stay open, the median and two further quantiles is explained in our article “Quantifying expert estimates in a Monte Carlo simulation: the Expert distribution”.

How do you assess the vulnerability?

The next factor is the vulnerability in row 6. It is important and relatively hard to assess. The vulnerability states what share of the attacks is expected to succeed. In the example, it is a constant value of 1%. If we have seen 200 attacks, we therefore expect two of them to be successful. This is not guaranteed; it remains random. For each individual attack in the year, you toss a coin that shows a successful attack with a probability of 1%, and then see how many successful attacks actually occurred.

In some cases, the vulnerability can be determined from historical data. There you can see how many attacks took place, how many updates, reboots and other events there were on the system, and how many problems occurred. This, too, can be corrected with expert judgement, because you must always assume a number of unreported cases. There are attacks that were not noticed at all and that turned out well because they did not penetrate the system. So here it is the other way round from the attack frequency: the attack frequency is potentially higher than you can see with the naked eye, whereas the vulnerability is lower, because there were blind attacks that were fended off without anyone noticing.

A second problem is that the strength of the attacks can change depending on which attacker gets their turn: a state-level hacker, or someone who is bored and just tries to see whether they can get in somewhere. The vulnerability can therefore change. An alternative to the constant vulnerability would be to make it random as well. This is not necessarily easier, though, because you do not know in advance within what range to make it random. So little is gained by it. It is possible, however, and it is done fairly often in practice.

How do you assess the direct and secondary losses?

For the third factor, the size of the direct losses in row 7, we also assume an Expert distribution. The same distribution can therefore be used on both sides, for the frequency as well as for the impact. The size of the losses also has a clear minimum of zero, because nobody routinely assumes that a cyber risk randomly generates income. The good case is T€ 350; the loss falls below it only with a probability of 1%. The median loss per successful attack is T€ 400. The loss stays below the bad case of T€ 600 with a probability of 99% and exceeds it only with a probability of 1%. Here, too, the maximum is left open. In rare cases the distribution can certainly take larger values.

If there is a successful attack at all, the question is whether secondary losses also arise, that is, whether something becomes public or there is an escalation beyond the normal level. In the example, we assume in row 8 that this happens with a probability of 90%. Estimating this probability is similarly difficult to estimating the vulnerability. It can change over time, and there can be quiet phases in which less happens. Here, too, you could assume a range and simulate the specific value of the probability. However, it is not obvious what this distribution should look like and which values to enter for it. The same therefore applies as for the vulnerability: a constant number that estimates the order of magnitude is actually enough.

The last input is the size of the secondary losses in row 9, again with an Expert distribution starting at zero. The good case is T€ 5,000, the median T€ 6,000 and the bad case T€ 8,500, with probabilities of 1% and 99% as for the direct losses. This represents the typical case in which the secondary losses are considerably larger than the direct ones. That is actually the normal case: setting something up again and getting it running costs relatively little compared with an escalation beyond the normal level.

In the example, the numbers have been chosen deliberately so that two moderate occurrences are larger than one bad one. Twice the median of T€ 6,000 gives T€ 12,000, more than the bad case of T€ 8,500. This way the evaluation clearly shows how often secondary losses occurred.

Can a secondary loss occur more than once?

In the example, we also assume that every successful attack can trigger secondary losses again. This can be a realistic approach, but it does not have to be. If, for example, confidential documents are leaked and made public, a repeat occurrence would not be an additional loss. If the same documents are published twice, no further damage arises; everything happened the first time. If, on the other hand, an attacker can obtain access codes for weapons systems, strategic information for military campaigns or information about the timing of certain activities again after some time through another attack, there are new losses as well.

How the secondary losses have to be handled must therefore be decided for the specific situation: either it is an event that can occur several times, or it is a Bernoulli event that can only occur once and is then over.

How do the factors work together in the model?

The middle block, “Rechnungen”, evaluates the inputs. Figure 4 shows it with the random numbers of a single run. At the top, the formula bar shows the formula for the number of attacks in cell C13. It follows directly from the Expert distribution with the parameters from row 5.

Figure 4: The calculations of a single run. At the top is the formula for the number of attacks in cell C13. In this run, 79 attacks lead to two successful attacks and two secondary losses, and the total losses amount to T€ 14,814 (German interface).

=RUNDEN(@Expert_variate(H5;C5;D5;E5;F5;G5);0)+@OutputName(B13)

The Expert_variate function draws a random number from the Expert distribution. Its arguments are the probability from H5, the minimum from C5, the good case, the median and the bad case from D5 to F5, and the maximum from G5, which is empty here. The Expert distribution also returns numbers with decimal places. The result is therefore rounded to a whole number with RUNDEN, the German Excel name of ROUND. If the distribution had drawn 79.35, for example, this would give the 79 attacks in Figure 4. The addition OutputName(B13) only gives the cell the name from B13 and does not change its value. The following formulas carry the same addition.

The frequency of successful attacks in cell C14 follows from the number of attacks in C13 and the vulnerability. The distribution that connects the two is the binomial distribution:

=+@WENN(C13=0;0;Binomial_variate(C6;C13))+@OutputName(B14)

For each of the attacks from C13, Binomial_variate tosses a coin that shows heads with the vulnerability from C6, that is, with a probability of 1%, and counts how often heads comes up. In the run in Figure 4 there are two successful attacks. The WENN condition in front of it (the German Excel name of IF) makes sure that zero attacks also give zero successful attacks.

In the third step, we determine the frequency of secondary losses in cell C15. The approach is exactly the same as for the successful attacks. The only difference is the basis: it is now the number of successful attacks from C14. Unsuccessful attacks do not cause secondary losses and play no role here. So for each of the two successful attacks we toss a coin that shows heads with the probability from C8, that is, with 90%. In this run heads comes up twice, so there are two secondary losses. The formula is built accordingly:

=+@WENN(C14=0;0;Binomial_variate(C8;C14))+@OutputName(B15)

More on this: Our article “When do I use a binomial distribution in a Monte Carlo simulation?” explains this distribution in more detail, including the case in which the number of trials is itself random, like the number of attacks here.

Now we come to the amounts of money. For the size of the direct losses in cell C16, one value has to be drawn from the distribution of the direct losses for each successful attack. The compound distribution from Figure 2 does this:

=@Compound(C14;"Expert";H7;C7;D7;E7;F7;G7)+@OutputName(B16)

The first argument, C14, states how often the event occurred at random. The second argument names the distribution from which the individual losses are drawn, here the Expert distribution. It is followed by its parameters from row 7 as a list. This general form, with the distribution as a name and the parameters as a list, is chosen because it allows any impact distribution to be used. It does not have to be Expert, and it can also have different parameters. In the run in Figure 4, the two successful attacks lead to direct losses of T€ 808 in total.

The size of the secondary losses in cell C17 is calculated in the same way with the compound distribution, from the frequency of secondary losses in C15 and the distribution for the size of the secondary losses in row 9. In this run it is T€ 14,006. Finally, the total losses in cell C18 are the sum of the direct and the secondary losses, here T€ 808 plus T€ 14,006, that is, T€ 14,814.

What does the simulation show for a single cyber risk?

Now we run the simulation. To do so, we click “Simulation” on the Risk Kit toolbar and calculate 100,000 simulation runs. Figure 5 shows the model afterwards. The “Rechnungen” block now shows a single run, here with 46 attacks and one successful attack. The “Ergebnisse” block now contains numbers. Their cells are formatted with the € sign only. As everywhere in the model, however, thousands of euros are meant, that is, T€. The same applies to the later tables that only show €.

Figure 5: The model after 100,000 simulation runs. The “Ergebnisse” block shows the probability of occurrence, the expected loss, the median, the values at risk and the maximum of the total losses. Cell C22 with the formula for the probability of occurrence is selected (German interface).

First we see the probability of occurrence of the risk in cell C22. It is 27.3%. It is measured with the Probability function, whose formula is shown at the top in the formula bar:

=+@Probability(RiskKit_Output;FALSCH;0)

Probability calculates the probability that the value in the output cell RiskKit_Output, that is, our total losses, is greater than zero. FALSCH, the German Excel name of FALSE, means: not smaller than or equal to, but greater than zero. If the total loss is greater than zero, the risk has occurred. Conversely, in 72.7% of the simulation runs there was no occurrence, and the total loss is zero there.

The expected loss, that is, the mean over all simulation runs, is T€ 1,949. The median is zero. This makes sense, because the probability of occurrence is only 27% and not above 50%. So in more than half of the runs nothing happens at all. Below this come the values at risk. The value at risk (VaR) is a quantile of the loss distribution: the VaR at 90% is the loss that is not exceeded in 90% of the simulation runs. Here it is T€ 6,808. The higher values at risk are correspondingly higher, up to T€ 21,801 for the VaR at 99.9%. The largest loss that occurred at all in the 100,000 simulation runs is T€ 40,379.

A hill of its own for each occurrence

The distribution of the total losses can also be viewed as a chart. Figure 6 shows it as a histogram. The horizontal axis shows the total losses in TEUR, that is, in thousands of euros, from 0 to 45,000. The vertical axis shows how many simulation runs fall into each loss range. You can see different clusters. The tall bar at zero stands mainly for the runs in which nothing happened at all. Further to the right follows the distribution of the secondary losses for one occurrence, two occurrences and so on.

Figure 6: Histogram of the total losses from the cyber risk across all 100,000 simulation runs. The tall bar on the far left stands mainly for the runs without a loss (German interface).

We are not really interested in the large bar at zero. So we zoom into the chart and enter the value 73 under “Min, %” in the window at the top right. The histogram then starts at the 73% quantile and only shows the range in which the risk occurred at all. Figure 7 shows the result.

Figure 7: The same histogram from the 73% quantile. On the left the narrow bar for runs with direct losses only, next to it one hill each for one, two, three and more secondary losses (German interface).

On the far left there is now a narrow bar. These are the runs in which the risk occurred but only direct and no secondary losses arose. Then follow a tall hill for one secondary loss and further, increasingly flat hills for two, three and more secondary losses. The distribution ends at about T€ 40,000, the maximum loss. That corresponds to about six secondary losses.

This risk therefore has a very exotic shape with a great many extreme values: there is a hill of its own for each occurrence, above all for each occurrence of the secondary event. An ordinary parametric distribution, that is, a distribution described by a formula with a few parameters, does not exist for this risk at all. Had we not simulated the distribution of the total losses, we would not be able to state its shape directly and would make very large errors in describing the impacts of this risk.

The functions for the results

Besides Probability for the probability of occurrence, the model uses the function Mean for the expected value, that is, the mean, the function MedianSim for the median and the function Perc for the quantiles in the “Ergebnisse” block. The maximum is effectively the 100% quantile; the function for it is simply called Maximum. Like Probability, all these functions evaluate the output cell RiskKit_Output, and Perc also uses the confidence level from column B.

The histogram was produced by the Plot function. Its formula is in cell B33, which is highlighted in orange in Figure 3:

=@Plot(C18;;"Gesamtverluste aus dem Cyber Risiko";"TEUR";;;"HISTOGRAM";1000;600)

The formula references cell C18 with the total losses, plus the title and the axis label TEUR. The numbers 1000 and 600 at the end are the width and height of the chart. They only matter if you want to embed the chart in the workbook. Then you also state the cell at which it should be embedded, and it appears there in this size. This was deliberately not done here because we wanted to zoom. Zooming is only possible in the live chart.

How do you combine many cyber risks?

There is a wide variety of cyber risks because you have to distinguish what feeds the attack frequency and what actually happens. This affects the shape and width of its distribution, the vulnerability and the size of the direct and secondary losses. The type of secondary losses also differs, because, as discussed above, they may only be able to occur once or several times. All these distinctions mean that you can end up with a huge number of cyber risks.

To be able to process the cyber risks further in enterprise-wide risk management, you have to rein in this variety again and consolidate the risk. This is best done by looking at many risks together and evaluating them together, in a risk inventory.

Figure 8 shows such a “Cyber-Risiko-Inventar” (cyber risk inventory). Each cyber risk has its own row, from Risiko-1 in row 6 to Risiko-30 in row 35. The columns contain the parameters we have seen in the model of the single risk: the attack frequency in columns C to I, the vulnerability in column J and, from column K, the size of the direct losses. Further to the right, outside the section shown, follow the probability and the size of the secondary losses. In the example, all 30 rows are identical copies of the same risk.

Figure 8: The cyber risk inventory with 30 cyber risks, one per row. It shows the parameters of the attack frequency, the vulnerability and the parameters for the size of the direct losses (German interface).

To the right follow the calculations in which the quantities are determined for each risk, and at the very end the total losses from each risk and the results. Figure 9 shows this part after the simulation. Column AE contains the total losses per risk in a single run. Here, “Gesamtverluste” (total losses) means the sum of the direct and the secondary losses of one risk. In row 7, for example, there are T€ 1,525 of direct losses in column AC and T€ 17,782 of secondary losses in column AD, T€ 19,307 in total. At the top right, in cell AP6, by contrast, is the “Gesamtverlust” (total loss) as the sum across all 30 cyber risks, T€ 73,744 in this run.

Figure 9: The risk inventory after the simulation. On the left the losses per risk in a single run, in the middle the results per risk, on the right the total loss across all 30 risks and its results (German interface).

After the simulation, columns AF to AN show the statistics of each individual risk. Essentially, all rows show the same values. That is only logical, since the 30 risks are identical copies. The probability of occurrence in column AF (EW) lies between 27.1% and 27.6% for all of them, the expected loss in column AG between T€ 1,931 and T€ 1,982. The median and the values at risk also differ only in nuances.

The only exception is the maximum in column AN. It takes clearly different values, ranging from T€ 34,561 in row 15 to T€ 66,822 in row 25. The reason is that this value does not stabilise when the number of simulation runs is increased. A loose end remains that swings back and forth like the lash of a whip.

With enough cyber risks, something is almost certain to happen

On the far right of Figure 9 are the results for all 30 risks together. The probability of occurrence is now 100.0%: something happened in practically every one of the 100,000 simulation runs. With enough cyber risks, something is therefore almost certain to happen. Risk management is therefore definitely needed at this level, and so is crisis management, because crises cannot be avoided 100%.

At the same time, the statistics have put things into perspective considerably. In the model of the single risk, the largest loss was about T€ 40,000 (Figure 5); for the 30 risks together it is now T€ 173,770. With 30 times as many risks, that is only a little more than four times as much. The portfolio therefore changes the picture in two directions: the probability that anything happens at all rises sharply to 100%, whereas the size of the losses rises clearly less than proportionally.

Figure 10 shows the distribution of the total loss from all cyber risks. The horizontal axis again shows the losses in TEUR, here from 0 to 200,000. The distribution again has a very exotic shape, like a mohawk with individual spikes. Because we essentially have the same risk 30 times with the same characteristics, the shape is nicely regular. The spikes come mainly from multiple occurrences of the secondary losses. You can count them: one secondary loss, two, three and so on up to about 14, after which the distribution tails off towards the top. These secondary losses can come from multiple occurrences of the same risk or from simultaneous occurrences of different risks.

Figure 10: Histogram of the total losses from all 30 cyber risks. The individual spikes stand for the number of secondary losses (German interface).

How do you transfer the result into enterprise risk management?

If you want to transfer all this into enterprise risk management, one question remains to be answered. This distribution has no standard shape. There is no parametric distribution that looks exactly like the distribution in Figure 10. So you need a way to transfer the results from the simulation into enterprise risk management without losing information.

This works by having Risk Kit output the simulation results. Risk Kit then writes the simulated total losses of all 100,000 runs into a separate worksheet, one value per run. This distribution can be transferred to the Enterprise Risk Explorer with a mouse click. It takes over the values as a fixed table of precomputed values, and the risk cluster is available in enterprise risk management for further analyses.

Can the approach be applied to other risks?

Let us summarise the approach. We started from an abstract cyber risk and broke it down. In other words, we carried out a disaggregation and looked at which individual factors influence the risk, what exactly they are, what shape they have and which parameters we need for them. Because they were tangible factors, we could in some cases also find data for them in the company. Then we aggregated the individual factors back into the risk.

With the portfolio approach, we went one step further and combined numerous cyber risks into a risk cluster, which we transferred into enterprise risk management. The result is an unbroken value chain that begins with breaking down a risk and ends with company-wide risk aggregation.

This basic approach can be used not only for cyber risks but for all kinds of risk types. In the first step, you break the risk down until everything abstract has disappeared and only tangible factors remain whose relationships you know. Then you aggregate them again in a Monte Carlo simulation and transfer the result for further use in the company.