RUNTIME: secureai runtime
Three enforcement engines for production systems: context access control, behavioural fingerprinting, and capability boundaries.
Unlike the earlier phases, RUNTIME evaluates a live session rather than source files. Each engine takes artefacts your application already produces, such as a context object, inference records, or a session, and decides whether what happened stayed inside policy.
LCAC. Context access control
Evaluates a context object against Policy-as-YAML across five pillars, reporting violations and redactions. Use it to confirm an agent only read the context it was entitled to.
secureai runtime lcac eval \
--policy payments-readonly.yaml \
--context context.json| Flag | Default | Description |
|---|---|---|
--policy | required | Policy-as-YAML file. |
--context | required | Context object JSON to evaluate. |
--format | terminal | Output format: terminal|json. |
Output lists each violation with its kind. scope for a read outside the granted boundary, capability for a tool call absent from the grant, plus the fields that were redacted.
LBF. Behaviour fingerprinting
Builds a behavioural baseline from normal traffic, then scores a live session against it. Drift from the baseline, such as eroding refusals, restated instructions, or a changing tone, appears as an anomaly score between 0 and 1.
1. Build a baseline
secureai runtime lbf baseline \
--records profiling-records.jsonl \
--out baseline.json| Flag | Default | Description |
|---|---|---|
--records | required | Profiling records, JSON or JSONL. |
--out | baseline.json | Where to write the baseline. |
2. Score a session
secureai runtime lbf score \
--baseline baseline.json \
--session session.jsonl \
--threshold 0.75| Flag | Default | Description |
|---|---|---|
--baseline | required | Baseline JSON produced by the baseline command. |
--session | required | Session inference records, JSON or JSONL. |
--threshold | 0.75 | Anomaly score above which the session is flagged. |
--format | terminal | Output format: terminal|json. |
Per-request output shows the anomaly score, whether it alerted, and the dominant signal. For example refusal erosion or instruction drift.
A baseline is only as good as its input
Build it from traffic you believe was clean. Baselining a period that already contained an ongoing attack teaches the model that the attack is normal, and the anomaly scores will look reassuring while nothing is being caught.
CBE. Capability boundary enforcement
Holds an agent inside a declared capability profile. When behaviour looks anomalous, the agent is degraded, for example to read-only, and any action outside the remaining boundary is recorded as a violation.
secureai runtime cbe eval \
--profile payments-agent.yaml \
--session session.json \
--anomaly 0.81| Flag | Default | Description |
|---|---|---|
--profile | required | Capability profile YAML. |
--session | required | Session JSON. |
--anomaly | latest from bus | LBF anomaly score, 0-1. Omit it and CBE reads the most recent score from the local Intelligence Bus. |
--format | terminal | Output format: terminal|json. |
Chaining the three
The engines are designed to feed each other: LBF produces the anomaly score, CBE consumes it to decide whether to degrade, and LCAC records what the session was allowed to touch. Run them in that order and CBE will pick up the score automatically.
secureai runtime lbf score --baseline baseline.json --session session.jsonl
secureai runtime cbe eval --profile payments-agent.yaml --session session.json
secureai runtime lcac eval --policy payments-readonly.yaml --context context.jsonWhere results land
- Anomalies and violations appear under Runtime events in the dashboard, visible to security researchers and org admins.
- Individual sessions are inspectable under Sessions, with the per-request anomaly trail.
See the Findings guide for triage.