Jorge Quitério Search
IT & Engineering

The role of information security in risk decisions

Security teams are often asked to approve or reject a risk. Decision theory explains why they should supply the probabilities and leave the choice to the risk owner, and it also indicates when further analysis is no longer worthwhile. A small model illustrates both points.

"Can we go live with this open finding?" Security managers regularly receive this question, often shortly before a planned launch.

There are two common answers. No positions security as an obstacle and encourages the business to bypass it. Yes means that security accepts risks it has no authority to accept and is held responsible when they materialise. Both answers are problematic.

This article argues that the role of security is to facilitate the decision rather than to make it, and shows that this follows directly from decision theory. Decision theory also addresses a question that arises in most risk meetings: should the decision be taken now, or should the analysis continue?

The relevant standards and literature are as follows.

The standards. ISO/IEC 27001:2022 requires that risk owners approve the risk treatment plan and accept the residual risk (clause 6.1.3). ISO/IEC 27005:2022 gives the guidance for doing so. NIST's Cybersecurity Framework 2.0 (2024) added a Govern function precisely to place risk strategy and accountability with leadership. ISACA's CISM body of knowledge frames the information security manager the same way: aligned to business objectives, advising rather than owning the risk decision.

Decision theory. The formal language is older. Raiffa and Schlaifer (1961) introduced the expected value of perfect information (EVPI), the most a decision maker should pay to remove uncertainty before choosing. Stochastic programming uses the same idea, as in my note on the value of the stochastic solution.

Why security should not own the decision

A risk decision balances a possible loss against a business objective: revenue, a deadline, a customer commitment, a mission. The person accountable for that objective is responsible for the trade-off. When security decides instead:

  • it becomes a blocker: every residual risk looks unacceptable to someone who does not own the upside; or
  • it becomes a rubber stamp: under pressure, it approves risks it has no authority to accept.

Beliefs and preferences

Decision theory splits any choice under uncertainty into two inputs:

  1. Beliefs: for each option aa, a probability distribution of the loss LaL_a. This is security's expertise.
  2. Preferences: how much expected loss, and how much tail risk, the organisation will carry for the upside. This belongs to the business.

Confusing the two produces both failure modes.

Medicine settled this division long ago. Before an operation, the surgeon explains the likely benefits, the risks and the alternatives; the patient, who will live with the outcome, decides. A surgeon who refuses to operate on any patient with a risk factor, or who operates without consent because the operation is usually successful, has confused expertise about probabilities with authority over preferences. Informed consent is the clinical form of a facilitated decision.

Structure of a facilitated decision

Facilitation does not mean neutrality. It means doing the analysis thoroughly and presenting it in terms the decision maker can act on:

Field Content
Objective The business outcome at stake, in the owner's words
Scenario What could go wrong, how, and to what
Assessment Likelihood and impact, stated in business terms
Options Mitigate, transfer, avoid or accept, each with cost and residual risk
Recommendation Security's view, clearly labelled as a recommendation
Decision The option chosen, by whom, and until when

Two details are important: impact in business terms (three days of downtime during a customer migration, not "CVSS 8.1"), and a review date, because a risk acceptance is valid for a defined period.

Escalation

If a proposed acceptance exceeds the organisation's risk appetite, or conflicts with a legal obligation, the risk owner cannot accept it alone. In that case the decision is escalated, with the same record, to the level that has the authority. That presupposes a written risk appetite; if none exists, writing one with senior leadership is often the most valuable thing a new security manager can do.

The expected value of perfect information

Risk meetings often conclude that further assessment is needed. Decision theory quantifies the value of that assessment. With θ\theta the unknown state of the world:

EVPI=min⁡aE[La]  −  E[min⁡aLa(θ)]. \text{EVPI} = \min_a \mathbb{E}[L_a] \;-\; \mathbb{E}\Big[\min_a L_a(\theta)\Big] .

The first term is the best expected loss achievable by deciding now; the second is the expected loss achievable if θ\theta were known before choosing. No assessment can be worth more than this difference. Clinicians apply the same test informally: a further examination is worth ordering only if its result could change the treatment.

The figure below applies this to a small security decision. EVPI is never negative, and when it is small compared with the stakes, further analysis is not worth its price.

The model is deliberately small. In a year when the organisation is actively targeted, the chance of a breach is 30%; otherwise it is 2%. A control removes 80% of the risk at an adjustable price.

  1. Close options. Move the cost of the control until "accept" and "add the control" are nearly equal. EVPI is usually largest near there: when the options are close, knowing the truth matters most.
  2. Clear decision. Set the loss to €10M. Adding the control is clearly preferable, EVPI shrinks relative to the cost, and the readout says "decide now". Further analysis would only confirm the decision.
  3. Non-negativity. No setting makes the "with perfect info" bar taller than the best decision.

Why EVPI ≥ 0. For every state θ\theta and every action aa, min⁡a′La′(θ)≤La(θ)\min_{a'} L_{a'}(\theta) \le L_a(\theta). Take expectations: E[min⁡a′La′(θ)]≤E[La(θ)]\mathbb{E}[\min_{a'} L_{a'}(\theta)] \le \mathbb{E}[L_a(\theta)] for every aa, and therefore also for the best one. Knowing more can never make the best achievable expected loss worse.

Security facilitates because a decision needs two inputs and security owns only one of them, the probabilities. The risk owner owns the preferences. Keeping the two apart makes the decision accountable and security's advice credible, and it leaves a record that auditors and future colleagues can follow. EVPI provides a useful test before further analysis is agreed: whether its result could change the decision by more than the analysis costs.

References

  1. ISO/IEC 27001:2022, Information security management systems - Requirements, clause 6.1.3.
  2. ISO/IEC 27005:2022, Guidance on managing information security risks.
  3. NIST, The NIST Cybersecurity Framework (CSF) 2.0, 2024.
  4. H. Raiffa and R. Schlaifer, Applied Statistical Decision Theory, Harvard University, 1961.
  5. ISACA, CISM Review Manual, 16th ed., 2022.