Risk Monitoring

How risk scoring works

Rather than one opaque number, Prefactor classifies risk on two axes. Data sensitivity reflects the regulatory and ethical weight of the information a run touches — personal identifiers, financial records, and health data each carry different implications if mishandled. Action consequence reflects what the agent did with that data — reading a record is different from writing one; retrieving information is different from initiating a transaction.

A risk profile assigned to an agent tells Prefactor how to weight each dimension — for example, reading data might carry a 1.0x multiplier, with other action types weighted differently — and those weights, multipliers, and thresholds combine into a classification: Low, Medium, High, or Critical. A High classification flags a run that involved high-sensitivity data, a high-consequence action, or both; it identifies a run that warrants closer attention, not proof that something went wrong.

Reviewing and acting on risk

Flagged runs land in a review workflow: a reviewer filters to a specific agent's instances, narrows to the environment and time window that matters, and opens the ones that reached a risk level worth attention — rather than scrolling every run or checking a scattered set of dashboards. Because the classification is attributed at the instance level and derived from its spans, a flagged run always breaks back down into the specific spans that drove the classification.

Where a team has runtime policies or approval routing configured, a flagged run can also route straight into those — throttle, sandbox, escalate to human-in-the-loop review, or block — rather than waiting for a manual pass.

Risk as a trend, not just a point-in-time score

A single run's classification tells you about that run. What it doesn't tell you is whether a given type of action — a specific tool call, a specific kind of data access — is becoming riskier over time, even if no individual instance crosses a threshold on its own. Prefactor tracks risk per span type the same way it tracks latency: as a distribution, with p95 and p99 views, so a slow drift toward riskier behaviour in one recurring action shows up as a trend before it shows up as an incident.