Using FAIR to Estimate Annual Loss Exposure from AI Incidents
To apply FAIR to AI incidents, define a narrow scenario (asset, threat community, threat type, effect), split it into threat event frequency, vulnerability, and primary and secondary loss, assign calibrated min, most likely, and max ranges, and run a Monte Carlo simulation to get an annual loss distribution. Then change the inputs that a control affects, mainly vulnerability and loss magnitude, and compare the results. With sparse AI incident history, wide ranges and documented assumptions are more defensible than precise numbers.
Scope one scenario at a time
Start with a narrow scenario. State the asset, threat community, threat type, and effect in one sentence, so that every later estimate refers to the same event.
Decompose the scenario into FAIR factors
Split the scenario into threat event frequency, vulnerability, and primary and secondary loss. Loss event frequency is threat event frequency multiplied by vulnerability. Loss magnitude is primary loss plus the probability-weighted secondary loss.
| Component | Description |
|---|---|
| Scenario | Asset, threat community, threat type, and effect, stated in one sentence. |
| Loss event frequency | Threat event frequency multiplied by vulnerability. |
| Loss magnitude | Primary loss plus the probability-weighted secondary loss. |
| Simulation | Monte Carlo draws from calibrated ranges produce an annual loss distribution. |
| Control comparison | Re-run with adjusted inputs to show the effect of a control change. |
Calibrate inputs when incident history is thin
Assign a calibrated min, most likely, and max range to each factor. Monte Carlo draws from these ranges produce an annual loss distribution. Because AI incident history is sparse, wide ranges and documented assumptions are more defensible than precise numbers.
Worked example with illustrative ranges
To model a control, re-run the simulation with adjusted inputs and compare the results. Change only the inputs that the control affects, mainly vulnerability and loss magnitude.
Map controls to FAIR factors
A control belongs in the model only through the factors it changes. Least-privilege permissions tied to a distinct agent identity raise resistance strength, because fewer manipulated actions succeed. They also cap loss magnitude, because the agent can reach fewer records. Tool-call policy enforcement raises resistance strength for each attempted call, whatever the model was persuaded to do. It lowers TEF only if it removes exposure to untrusted inputs, which is uncommon.
Audit logging does not prevent anything. Its effect is on detection and scoping: faster identification of what was accessed can reduce response cost and may lower the probability or size of secondary loss. In the example, that is why primary and secondary ranges shift down modestly while vulnerability shifts more.
Interpret results and avoid common errors
Compare the annual loss distributions from the baseline and adjusted runs, and keep the assumptions behind each range documented.
Support your estimates with enforceable controls and audit evidence
Runtime policy enforcement, agent identity, and audit logging can give risk teams observable data for vulnerability and loss inputs. Talk with Trussed AI about how these controls fit your scenarios.
Talk to an Expert