How to recruit the perfect employee (with free resources)

You may have arrived at this page thinking, “Okay, this is a clickbait article, but I’ll bite.” Clickbaity, because you’d say there’s no such thing as a perfect employee. And I confess, I’d half agree with you. That said, when hiring new people for a role, we’re usually looking not simply for someone who can just do the work but who also will be a good fit. When you get someone with both attributes, that’s pretty close to perfect, I’d say.

Interview processes are artificial. If like me you don’t have lots of time to spare for recruiting and you don’t have a department dedicated to taking new recruits through a fortnight-long getting-to-know-you exercise, you’ll appreciate some effective time-savers.

Most interviews I’ve conducted have been on the telephone or face-to-face in the office – and relatively short. These are not ideal scenarios for really getting to know someone. And for me, it is vitally important to know who the candidate is, not just what the candidate can do.

Will candidates fit in well with the team? Will they share our values? Will they bring toxicity and poor attitudes? Some people pride themselves on being a good judge of character. If this is you, I’m really, genuinely happy for you. The rest of us though, we need some tools to help us. Some tools that take us closer to understanding the core of someone’s being, within the time constraints of an hour-long interview.

Virtues toolkit

If you want to know what someone is like, one approach is to ask them to comment on their own character. Many people find that difficult enough when there’s no pressure, never mind in an interview situation, not least because they’re second-guessing, wondering what might be the ‘right answer’.

Of course there is no right answer. You will always and only be you. Even if you’re a good actor and can play the part of someone with a different personality, in time, under stress, the real you will emerge. And in the meantime, you’ll waste a lot of emotional energy.

So with our candidates, to get to the real person we need to draw out the values that are important to them. To what characteristics would they say they aspire? Chances are, they are already innately keyed into those characteristics and are striving to improve. Maybe they’re passionate about justice, or desperate to be more professionally dispassionate (to become a better negotiator). None of these characteristics or values are good or bad, per se. But some will be more pertinent to your business and in interactions within the team.

My virtues toolkit is really simple. Without judgement, I created a diverse list of characteristics that might be considered to be virtues. I printed them, laminated them and cut them up into individual pieces. Then within interviews I gave candidates the pile of virtues and asked them to select the five to which they most aspired or which they most admired.

Can candidates ‘game’ this system? Well yes, of course. But that’s why we have probationary periods! If you discover your candidates do not in any way reflect the values they claimed to espouse, you can be sure you have an issue of integrity, which may make them unsuitable to work in your business. Or possibly, if the gap is not too wide, it identifies areas for improvement during the probationary period or beyond.

Here are the templates. Feel free to use as they are or customise them according to your own tastes and requirements. At the very least I have no doubt you can make them look better. My design skills are feeble at best!

Categorisation exercise

This one’s a bit more specialised. It may or may not be of use to you, but if nothing else, I hope it might inspire other creative methods of conducting candidate assessments.

When I was recruiting for a Security Analyst position recently, I spent some time thinking about the qualities that might be advantageous for such a role. How might an analyst think?

In many different analyst roles, not only within information security, it is fundamentally important to be able to sift through data. To see the wood for the trees. To identify different characteristics in the information presented. To spot those factors that are relevant. To think like an analyst, you probably need to be pretty good at classification, categorisation, developing or utilising taxonomies.

As I said before, finding the right person for the role is important – possibly more important than finding someone with the right experience. If you recruit someone who has all the right tendencies and the ability to learn quickly, a lack of experience is of less importance in many roles. (Leaving aside for the moment those roles where prestige and documented credibility or political power are key.)

So I developed a categorisation exercise. I produced a sheet of 48 different items of clothing, all different from each other in various ways. Again, I laminated them and cut them up into individual cards. I gave these to the candidates and invited them to make notes on the various different ways in which you might categorise the cards – and whether a particular method stood out as most useful or appropriate.

While the candidates completed the task, I observed and made notes on how readily they adapted to this and how quickly they were able to work. A logical analytical mind makes light work of this sort of task, even if the person has no previous experience of a game like this. (Let’s face it. It’s definitely a game, albeit not a very fun game.)

I made it clear to the candidates that this was not a colour perception test. The results were illuminating and as far as I can tell, their interaction with this test correlated strongly with what I knew about them or later came to know.

The original template for this exercise is in SVG format. You can edit this using the free (and very competent) graphics program Inkscape.

Conclusion

With a bit of innovation and some lateral thinking, there are definitely some strategies you can use to get closer to understanding the people you interview. Clearly there’s no substitute for putting your candidates at their ease, giving them time and space and getting to know them. But for a busy hiring manager, these tips and resources might get you a shade closer to perfect – and if not these exact resources, whatever new resources you are now inspired to create.

I’d love to hear if this post has helped you or if you have any other ideas about how we can improve the interview process for recruiter and candidate. Let me know your thoughts in the comments below. If you have great resources of your own, get in touch and perhaps I can provide links to them here.

Better than a risk matrix

There’s something better than a risk matrix? It’s a bold claim. But risk matrices have significant weaknesses, as I have discussed elsewhere.

In information security, we know (I hope) that our role is primarily concerned with the control of risk. We may agree that’s what we’re doing – but unless we can measure risk and show how our efforts change our risk exposure, where’s our credibility?

When suggesting we should measure risk in infosec, a common objection is that we can’t measure this – there are too many unknowns. But we do ourselves a disservice here: we know a lot. We’ve trained, we have experience. And much as I hate to use the term, we’re experts. (Certainly when compared to people outside our field.)

So we take our expert opinions, maybe some historical data and we feed it into a model. ‘Modelling’ can be an intimidating concept until we understand what we mean. We model all the time – tabletop DR exercises, threat modelling, run books, KPI projections. The point is once we know what we’re doing, it’s not intimidating at all.

So let’s get over that hump.

Risk modelling is a matter of statistics. Again, don’t let that put you off. You don’t need to be a stats whizz to use a statistical model. Many of us are engineers. We know how to use tools; we like using tools; we often make our own tools. Statistical models are simply tools; we just need to learn how to use them. We don’t necessarily need to be able to derive them. But usually we like to have some idea of what’s going on ‘under the hood’. So here’s a whistle-stop tour of the concepts underpinning the model I present below.

Distributions

Let’s say we’re finding out how many televisions there are in the households in our city. We take a poll, where each household reports how many televisions they have and then we plot a graph. The x-axis shows number of televisions, say from one to ten. The y-axis shows number of households with that number of televisions.

We might expect to see a graph a bit like this (yeah, that’s a lot of televisions):

This is a normal distribution. Normal distributions are common in statistics and produce a curve that’s often called a ‘bell curve’, due to its shape. The data tends to congregate around a central point, with roughly equal values either side. The amount of spread is called the ‘standard deviation’.

Reality often shapes itself into a normal distribution. But in information security risk, you may find that other types of distribution better reflect the circumstances. One such option is the log-normal distribution. Here’s its curve:

If we’re measuring the impact of an event, a log-normal distribution often fits the bill. It does not go below zero (an impact by definition means an above-zero loss). The values tend to congregate toward the left of the graph, but it leaves open the option for a low probability of a very high number, to the right of the graph.

What we’re doing is creating a mathematical/statistical model of reality. Since it’s a model, we choose whatever tool works best. A log-normal distribution is a good starting point.

Confidence interval

In measuring risk, we take account of our uncertainty. If we were in the enviable position of having absolute certainty about future events, we would be better off turning this astonishing talent to gambling. As it is, we’re not sure about the likelihood and impact of detrimental events, and hence we estimate.

When estimating impact, we could simply select a value from one to five and produce a risk matrix. But as we saw before, this is not particularly informative and in fact it can be misleading. One aspect that the ordinal scoring overlooks is that impact is best expressed as a range. If the negative event occurs, e.g. we experience a ransomware outbreak, that may prove to be a low-impact incident, or things may go horribly wrong and it costs us millions.

The confidence interval allows us to express a range for the impact, based on our level of uncertainty.

A confidence interval of 90% is often used. This means that the actual impact has a 5% chance of being above this range and a 5% chance of being below the range. In practice this proves to be good enough for our purposes.

Side-note: the log-normal curve that models a 90% confidence interval has a standard deviation of 3.29. We’ll use this fact shortly.

Imagine you have ten identical slips of paper. On nine of them, you write ‘winner’; on the other one, you write ‘loser’. All ten slips are placed into a hat and you draw one out at random. If you draw the ‘winner’ slip, you win £1,000. Otherwise, you win nothing.

On the other hand, there is a football match coming up today and your favourite team is playing. You are asked to predict the likely number of goals your team will score, within a range, with 90% confidence (nine in ten). Again, if you are right, you win £1,000.

In the book How to Measure Anything in Cybersecurity Risk, the authors call this an ‘equivalent bet’ test. If you prefer the idea of drawing slips out of a hat, because you think you’re more likely to win, that means you didn’t really use a 90% confidence interval for the football score prediction. You need to widen the range.

On the other hand, if you prefer the football bet, that means that your range prediction was probably too wide. You’ve used perhaps a 95% confidence interval.

The trick is to balance the two such that the potential reward is equivalent in either scenario. In this way, you will have achieved a true 90% C.I. with your football prediction. It takes a little effort to wrap your head around this but press on: this is an invaluable concept in risk analysis, which we’ll use shortly.

Estimating likelihood

When you’ve defined your threat event, estimating its likelihood is straightforward: define a time period and a probability (percentage likelihood) of the event occurring. To be meaningful, don’t make the event too specific. So ‘a ransomware outbreak’ is probably a better event definition for most companies than ‘a solar flare causing communications anomalies that disrupt global networking such that intercontinental backups are delayed by three hours’.

Over a 12-month period, you may estimate that you have a 5% likelihood of experiencing a ransomware outbreak. This percentage need be no more than the considered opinion of one or more experts. You can improve the estimation through the use of data: historical ransomware attacks in your sector, global threat activity, etc. But the data is no more required for this model than it was for a traditional risk matrix.

At this stage, do not consider the severity of the attack, just the likelihood. This likelihood percentage stands in the place of the 1 to 5 numbering in the risk matrix; it is simply more helpful.

Estimating impact

For impact, you now estimate a range, using a 90% confidence interval. You can use the equivalent bet test if you like, to guide your estimate. Again, you can take into account any available data, including known costs of breaches as they are reported worldwide. You can also consider things like the possible duration of an outage, the costs associated to an outage, the costs of paying the ransom or the excess on a cyber risk insurance policy. So you might say, all things considered, you are 90% confident that the impact would be between £250k and £1.5m.

We’re going to use a log-normal distribution, as shown above. During modelling, due to the long tail on this graph, you may find that the model produces some extremely high values that are simply not realistic (£75m, say) – unrealistic for whatever reason, such as the fact perhaps this exceeds your company’s annual global sales. You can therefore introduce a cap to this impact; e.g. a 90% C.I. of £250k to £1.5m, capped at £10m.

Modelling the risk

You now have all the data you need, to model the risk. What does this mean? Merely that you will insert these numbers into a formula and find out what happens when you run the formula a thousand, ten thousand, a hundred thousand times.

And why will you see different answers each time you apply the formula? Because we introduce two random variables. The random variables are based on all the above. The first random variable represents whether the event occurred ‘this year’, based the likelihood probability. To express this as a formula in Excel, Google Sheets or >insert favourite spreadsheet editor here< you do:

= IF (RAND() > likelihood, 0, 1)

Here, likelihood is the percentage expressed as a decimal (0.05). And RAND() produces a decimal number less than 1. So if RAND() is greater than 5%, the event didn’t occur and the result is 0. If RAND() is less than 5%, it did occur, hence the result is 1. We will multiply this by the impact, which we calculate next.

For impact, we use an inverse log-normal function. We chose to use a log-normal distribution above, we know the impact is somewhere within that distribution and again we use RAND() to work out the precise impact (‘this time’) based on this knowledge. Remember that with a 90% CI, there’s a 5% chance the impact will be higher or lower than the range we specified. And remember further that 3.29 is the standard deviation for such a curve.

So we have the range and the confidence interval and we’re working back to a single figure. The formula is this:

= LOGNORMAL.INV(RAND(), mean, standard deviation)

Or, where high and low represent the upper and lower bounds:

= LOGNORMAL.INV(
  RAND(), 
  ( LN( high ) + LN( low ) ) / 2,
  ( LN( high ) - LN( low ) ) / 3.29
)

Use LOGNORMAL.INV in Excel. In Google Sheets the function is LOGNORM.INV. With log-normal distributions, the figures are based on the log of the mean and standard deviation – that’s why you see here the log function LN.

To prove that you don’t need fancy software or a powerful computer to do this modelling, I ran this simulation using Google Sheets on a modestly-specced Chromebook, over 1,000 rounds. You can download the spreadsheet at the end of this article to see how this worked. To cut a long story short, on one run of this simulation I ended up with the figure £45,540.19, being the annual risk exposure for this threat. (With just 1,000 rounds, you can expect to see some variation each time you recalculate, but it’s enough to demonstrate the model.)

You’ll see in the spreadsheet some columns mysteriously labelled “Loss Exceedance Calculations”. At some point, I may write an article to address that!

For now, I hope this has whetted your appetite and given you enough to start improving upon your risk matrices. Happy modelling! And if you’re interested in learning a bit more about statistics in general (some measure of stats fluency is well worth it), I suggest taking a look at Statistics for Dummies. Don’t be put off by the title!

How a risk matrix can kill you

As a security professional, unless you work for an MSSP (Managed Security Service Provider), security is simply a cost to the business. Fitting a swipe card door entry system will not in itself generate more revenue. Increased password complexity rules are unlikely to result in increased sales.

How then do we justify our existence? By the way we reduce risk.

If you work in penetration testing or you’re a network security engineer, you might find this to be a very unsexy subject. Risk. Boring. But it underpins everything you do and should be at the heart of every choice you make. We do not secure our networks simply because the manual says we should, or because the compliance department insists on it; we do it because it reduces risk. In reducing risk, we’re improving stability, longevity, our ability to service our customers; in short, we’re protecting our jobs.

Well then. How do you know if your actions as a pen tester or engineer reduce risk? Through risk assessments. Stay with me.

You may already have been subject to a risk assessment in your workplace – checking you know how to use your office chair, for example. And if this is your only experience of risk assessment, I can understand why you might be completely put off by the whole thing.

But there’s more to it than that. And risk assessment in information security can be a lot more interesting (and less patronising). Truly. I’m an engineer by disposition and yet I’m excited about risk.

Now if you’ve been around for any length of time, you’ve no doubt seen one of these, the humble risk matrix:

5 x 5 risk matrix

On the face of this, a risk matrix couldn’t be easier to understand. (Perhaps that’s why we like them.) For any given risk, figure out the likelihood of it occurring (on a scale from one to five) and also the consequence, again on a scale of one to five. You can now position this risk on the matrix. And hey, maybe even multiply those two numbers together, to give you a single overall risk figure. Why not. Likelihood 4, consequence 5, okay that’s a risk of 20. Simple.

But what do we mean by ’20’? What do you mean? What does your CEO mean? Is it an absolute number? Or relative? Is a 20 twice as bad as a 10? The fact is that different people can have completely different perspectives on the meaning of these figures. We only have the illusion of communication. We’re using ordinal numbers, not cardinals.

A five by five risk matrix is an example of qualitative (as opposed to quantitative) risk analysis. A relatively arbitrary value is assigned on the basis of a subjective view of a risk.

Some organisations partially overcome the limitations of this approach by assigning explicit values to different ranks in the risk matrix. The likelihood can be split into five bands of 20% each, for example. Financial values are applied to impact. This is semi-quantitative analysis. E.g.:

Such divisions can also be misleading, unfortunately. Consider:

  1. A 41% likelihood (3) of a £5.1m impact (4) would be scored as 12.
  2. A 60% likelihood (3) of a £5m impact (3) would be scored as 9.

In this risk matrix analysis, risk a seems to be more severe than risk b. But consider the probability calculations:

  1. A 41% likelihood of a £5.1m impact gives financial exposure of £2,091,000.
  2. A 60% likelihood of a £5m impact gives financial exposure of £3,000,000.

So the risk matrix shows a to be the higher risk, but the calculation shows b to be higher. Both can’t be correct. Using a probability scale combined with likelihood range slices demonstrably results in incorrect prioritisation. (And a lack of these clarifying measures must inevitably result in even less reliable ‘calculations’.)

This may seem like a trivial example, but this could be the difference between a project being approved or postponed – on the basis of faulty analysis. And if that project is life-or-death – which it might be, in military or healthcare applications – then the risk matrix could indeed kill you*.

I’ll leave it on that bombshell – in later posts I will discuss some alternatives.

*(You got me, I’m exaggerating for effect.)