Risk Assessment
Every agent, workflow, and MCP client gets a composite risk score. The score is not cosmetic — it drives which policies apply, how long scopes stay valid, and how closely runs are sampled into the audit pipeline.
Four dimensions
Stryda scores four dimensions independently and composites them into one 0 – 100 number. The dimensions stay visible in the UI so you can see why a system is flagged, not just that it is.
Bias risk
Privacy risk
Operational risk
Regulatory risk
How the score is computed
Each dimension contributes a 0 – 100 sub-score based on signals we collect automatically:
- Scope surface — namespaces and tools the agent is authorized to use.
- Data classification — labels on the resources each tool touches (PII, financial, health).
- Autonomy — how many tools run without human approval in a typical week.
- Regulatory tags — frameworks the workspace has enabled and which of them apply to this system's surface.
The composite score is a weighted mean, with weights configurable per workspace. Systems above your threshold (default: 60) are flagged in the registry and get stricter policy defaults — mandatory approvals on sensitive tools, shorter scope lifetimes, denser audit sampling.
Runtime anomaly detection
Assessment is not just pre-execution. While a run is in flight, the pipeline watches for anomalous behavior — sudden spikes in cost, tools being called that were never called before by this agent, prompts that diverge sharply from the normal distribution. An anomaly raises the effective risk score for the run, which can trigger a pause even if the static policy would have allowed it.
When assessments run
- On registration — a new agent or MCP client is assessed the first time it calls
/api/mcp. - On scope change — widening a scope re-runs the assessment for every agent it touches.
- Nightly — a recomputation job picks up data-classification changes and regulatory-tag updates.
- Manually — run a fresh assessment from the Risk page for any system.