Install a Policy Pack
A policy pack is a curated set of CEL policies for one governance area, ready to install as-is or tune to your fleet. For how policies work in general, see Policies.
Available packs
| Pack | What it governs |
|---|---|
| Anomaly Detection | Session-anomaly counters and behavioral-baseline deviation — hard stops, HITL escalation, and audit-only signals. |
| Network Egress | Outbound destination allow-listing, SSRF detection, and command-and-control traffic blocking. |
| Approvals | Human-in-the-loop routing for production deploys and privileged autonomous-agent operations. |
| Content Safety | PII, credential, and unsafe-content scanning on requests and tool responses, plus prompt-injection detection. |
| Cost Guardrails | Per-turn cost efficiency, session age, context size, rate-limit saturation, and budget-threshold warnings. |
| Model Access | Model allowlisting, tier-based approval, group/team/token model scoping, and data-residency requirements. |
| MCP Tool Access | MCP tool-name allowlisting, sensitive-path and secret-file protection, shell-command allowlisting, and production namespace scoping. |
Every policy in every pack ships with a conservative default — most start as Audit (observe only) or Warn/Approve rather than Deny, so installing a pack never silently blocks existing traffic.
Install a pack
- From the UI
- At self-hosted setup
- Open Policies, then Add Policy.
- Paste in the YAML for each policy you want from the pack. Each policy in a pack is a complete, independent
ValidatingPolicydocument — install one, a few, or all of them. - Leave new policies in their shipped enforcement mode (
AuditorWarn/Approve) until you've reviewed matches for your workload, then graduate toDeny— see Write a Policy for the general Audit → Warn → Deny workflow.
If you're initializing a new self-hosted instance from the command line, install one or more packs in the same step as your config:
aicontrols init --packs=network-egress,approvals,content-safety
Combine --packs with --preset when you're also running a guided setup for a specific integration. Installing a pack is additive — it never removes or overwrites a policy you didn't get from that pack, and re-running init with the same --packs value is safe.
Tuning dispositions
Every pack's policies use the same enforcement modes as the rest of the product — Audit, Warn, Warn combined with a developer confirmation prompt, Deny, or Approve (human-in-the-loop). Review matches with Replay before promoting any policy past Audit, and tune the numeric thresholds in each policy's expression (percentages, token counts, dollar amounts, time windows) to match your fleet's actual behavior.
The Anomaly Detection pack currently reacts to session-anomaly counters and baseline-deviation state only. Fuller behavioral-detector coverage is planned as a future update to the pack — treat it as a starting point today, not a complete replacement for a dedicated anomaly-detection engine.