This is something that stuck in my head while at work today.
WARNING: NOT NECESSARILY ABOUT GAMES
The general case
Talents that are new to humanity go through four phases. Well, on different axes they go through all kinds of phases, but there’s one progression I’m interested in today.
Mysticism. At first there are few people with the talent and it is largely unexamined. Even the practitioners don’t really know how they do what they do. They have talent and inspiration and they seem to be effective. There are individual heroes and we tolerate a lot of bullshit because there’s not much out there but heroes at this stage. The word “genius” gets thrown around a lot.
Organized Mysticism. Once our mystics recognize that they have something special they organize. The find other mystics and grant them access to the organization. They deny access to those that don’t have it. This may or may not be literally organized, but there’s at least a social aggregation.
Investigation. At some point people realize that there can’t be anything magical or purely intuitive about this. There must be a way that people with the talent do what they do. Something we can quantify and proceduralize. This requires an honest and rigorous analysis of the talent and the talented.
Engineering. Once the talent is quantified we can teach it to others. No longer do we rely on the intuitive talent of individuals nor (in some cases worse) the accreditation of an individual by a mystic cabal. It can be taught and it can be tested and it can be reproduced. Anyone who wants this talent can have it.
One problem that arises is that during the Organized Mysticism phase there will be a lot of resistance to investigation. There is significant pressure to remain mystical!
First it’s a lot less work because people can only check your results and not your process. And your results don’t have to be all that good to be good enough — just a little better than a random guess. In reality you don’t even need to be that good if your successes are spectacular enough or the failures of those who don’t use your mystic organization are publicized properly.
Second it’s lucrative. You control access to the talent, so you can price it however you like. And then you also control membership to the Mystic Cabal and if your outcomes aren’t all that controlled, maybe you just want to sell some memberships and make a packet that way. This may or may not happen but the pressure is there and the controls are absent.
And investigation is expensive and has no immediate pay off. It’s an academic exercise, one done for the love of the knowledge. It’s a future-value endeavour and one that may or may not pay off. I mean, we might discover that the talent doesn’t actually exist and then you are stuck at the Organized Mysticism stage and you are discredited. The value in self examination is low.
And honestly if you have an amazing intuitive talent do you really want to be surrounded next year by people — just anyone really — doing what you do? That’s bound to bring down salaries.
So getting out of the Organized Mysticism phase is hard. It’s an ethical move. It should be the next step for any mystics who honestly believe that their talent is both valuable (to humanity — being valuable to yourself is actually a negative motivator here) and real. Resistance to investigation is suspicious.
The specific case
In the standard way of doing security risk assessments there is this idea of a risk calculation matrix, in which you cross index the impact of an event with the likelihood of an event to determine just how bad a threat is and therefore how much you should spend to mitigate it. At its root this is a good idea — it comes from safety analysis, after all, which is a time honoured science.
However, what we do here in co-opting this mechanism for security is not science, and it’s very much to our advantage as “experts” (especially certified experts) for it not to become a science. As long as it’s an art we don’t have to do much real work and at the same time our job seems like it’s a lot more clever than it is.
In a safety case, since we are dealing with an event tree that triggers on equipment failure (that is, on mean time to fail numbers — published numbers) rather than malicious activity, that “frequency” or even less credibly “probability” column is an actual number you get from a manufacturer. My fault tree shows that if component A and component B fail simultaneously then I cannot guarantee the system is safe. A and B both have published mean time between failure numbers (which are both measured and very conservative). The probability column here is just arithmetic.
In a security case that probability column is a Wild Assed Guess. We cloak it in two things: our credentialed “expertise” and by refusing to assert real (and therefore unsupportable since there are no real) numbers but rather vague order-of-magnitude categories. A first glance at the problem might suggest that this is just inevitable — the probability of malicious activity is not quantifiable. To me, though, this should not imply that we simply trust the instinct of a credentialled expert to suddenly make it quantifiable because the problem isn’t that it’s hard to know and that you need a lot of training and experience to estimate it. The problem is that it’s genuinely unknowable. That means when someone tells you they can quantify it, even vaguely, at an order-of-magnitude level, they are lying to you.
Unfortunately this lie is part of the training. You even get tested on it.
This makes us a (currently powerful) cabal of mystics. And the problem with a cabal of mystics being in charge is that first, they aren’t helping because they are not doing any science and second, as soon as someone starts doing some science they will entirely evaporate, exposed as charlatans. So naturally for those invested in the mysticism there will be some resistance to improving the situation.
The essence of science, setting aside for a moment the logical process (and that’s a big ask but it’s out of scope here) is measurement.
One axis of that risk calculation matrix is measured. The impact. Now it might be measured vaguely, but you can go down the list of items that qualify an event for an impact category and agree that the event belongs there. Someone could get seriously injured. Tick. Someone could get killed? Nope. Okay it goes in the SIGNIFICANT column. It’s lightweight as measurement goes but it’s good enough and it’s mechanizable (and that’s a red flag that separates engineers from mystics). You don’t need a vaguely defined expertise to be able to judge this. Anyone can do it if they understand the context and the concepts.
So the question I keep banging my head against is the other axis: frequency or probability. And since this is both unmeasurable and also has vast error bars (presumably to somehow account for the unmeasurability, but honestly if it’s impossible to measure then the error bars should be infinite — an order of magnitude is just painting a broken fence) my opinion is that it should be discarded. Sure it’s familiar because of safety analysis, but they have an axis they can measure. This one is not measurable. It’s therefore the wrong axis.
A plausible (and at least estimable if not measurable) axis is cost to effect. How much does it cost to execute the attack? This has a number of advantages:
- You can estimate it and you can back up your estimate with some logic. There’s a time component, a risk of incarceration, expertise, and some other factors. You can break it down and make an estimate that’s not entirely ad hoc and is better than an order of magnitude.
- It reveals multiple mitigations when examined in detail.
- It reveals information about the opposition. Actors with billions to spend might not be on your radar for policy reasons. Threats that can be realized for the cost of a cup of coffee cannot be ignored — you can hardly be said to be doing due diligence if attacking the system is that cheap.
- It is easily re-estimated over time because you retain the logic by which you established the costs. When you re-do the assessment in a year’s time and a component that cost a million dollars now costs a hundred, the change in the threat is reflected automatically in the matrix. No new magic wand needs to be waved. It’s starting to feel sciencey.
A useful cost to attack estimate (and I have nothing against estimates, I just expect them to be defensible and quantified) would need some standardized elements. For example, I would want us to largely agree on what the cost is of a threat of imprisonment. If I wet my finger and wave it in the air I’m happy with a hundred grand per year (a fair salary) of likely incarceration times about 10% for chance of getting caught. If we’re not happy with the estimate we can do some research and find our what the chances of getting caught really are and what the sentencing is like. We might find out that I’m being way too expensive here.
This is a good sign though. When I am compelled to say “we ought to do some research” I am happily thinking that we are getting closer to a science. What credible research could you do on probability of attack? Where would you even begin? And what would its window of value be? Or its geographic dependencies? Or its dependencies on the type of business the customer does?
Because you want to break the cost to attack down into the various costs imposed on the attacker — their time, their risk, their equipment costs — you have grounds to undermine the attack with individual mitigations. What if a fast attack took many hours? What if you could substantially increase the chance of catching them? What if you could increase the chance of incarcerating them? Suddenly those legal burdens start looking like they could be doing you a favour: you make this attack less likely by increasing your ability to gather evidence and to work with law enforcement. Publish it. Make an actual case and win it. Your risk goes down. These are mitigations that are underexplored by the current model but that could do some genuine good for the entire landscape if taken seriously. Sadly they don’t imply flashy new technologies at fifty grand a crack. But I am not interested in selling you anything. I want your security to improve.
In most of our assessments the threat vector, the person attacking, is categorized fairly uselessly into “hacker” and “terrorist” and “criminal” and so on. But their motivation doesn’t actually help you all that much. This isn’t useful information. How much they are willing to spend, however, does tell you about them. It tells you plenty. If you have a policy that you are only interested in threats from below a government level, that is that you aren’t taking action to protect yourself from a hostile nation state (and this is perfectly reasonable since it’s probably insurable: check your policies) then what you really want to do is decide how much money gets spent by an attacker before they qualify as a nation state? As organized crime? As industrial espionage? And so on? If you can put dollars to these categories then you can not only make intelligent decisions about mitigations but those decisions and the arguments behind them might even have some weight with your insurance adjuster. That’d be nice.
Finally these threats all change over time. Legislation changes, law enforcement focus changes, technology changes. But all of these changes are reflected in some component of the cost to attack. Consequently the value is possible to re-assess regularly. A vague value with no measurements is harder to justify re-considering — the whole thing starts to unravel if you ever wonder whether or not it’s right. Because it has no fabric to begin with. It’s just smoke and mirrors. It’s better not to look behind the curtain in that case.
But it’s much better to build on a foundation of measurement. It’s always better to have a calculation that you can expose to reasoned debate than to shrug and trust an “expert”. None of this is so complicated that no one can understand it without training. Making it seem so is a threat to doing the job properly. Let’s throw back the curtain and make this a science again. Let’s measure things.