Policy Templates

Versioned enforcement templates with a controlled lifecycle. Rules are append-only — changes supersede, never overwrite, so evidence stays reproducible.

LifecycleDraftIn reviewApprovedDeployedSupersededEvery change is dry-run against live traffic before deploy; rollback re-activates the previous version.

Financial Strict

v2.3 · updated 2026-06-24

Deployed

Hard blocks on PCI/PII exposure, constrain on internal identifiers, prompt-injection zero-tolerance.

PIIData leakagePrompt injectionToxicity

46 rules2 tenants assignedpolicy-team

Healthcare PHI

v1.9 · updated 2026-06-18

Deployed

PHI-first template: strict masking of patient identifiers, ingest-side document screening.

PIIData leakageToxicity

38 rules1 tenant assignedpolicy-team

Public Sector

v3.1 · updated 2026-06-12

Deployed

Arabic/English dual-language rules, national-ID masking, records-retention aligned evidence.

PIIData leakagePrompt injection

41 rules1 tenant assignedgov-desk

Baseline

v4.0 · updated 2026-05-30

Deployed

Default protection tier for standard commercial tenants. Allow-first with PII constrain.

PIIToxicity

24 rules2 tenants assignedpolicy-team

Financial Strict

v2.4-rc1 · updated 2026-07-01

In review

Adds jailbreak-chain detection rules and response-side constrain for account numbers.

PIIData leakagePrompt injectionToxicity

49 rules0 tenants assignedpolicy-team

Legacy Baseline

v1.7 · updated 2025-11-02

Superseded

Superseded by Baseline v4.0 — kept for audit reproducibility only.

PII

18 rules0 tenants assignedpolicy-team

Kept for audit reproducibility.

Baseline

v4.1 · updated 2026-06-28

Approved

Adds jailbreak-chain coverage to the default tier. Approved — awaiting deployment window.

PIIToxicityPrompt injection

26 rules0 tenants assignedpolicy-team

Rule semantics

How a template decides — mirrors the gateway's enforcement engine

Risk category	Severity	Outcome	Behaviour at the gateway
PII	high	block	Request rejected; customer-safe refusal returned; evidence recorded.
PII	medium	constrain	Detected spans masked as [pii] before forwarding upstream.
Data leakage	high	block	Blocked on both prompt and response sides of the exchange.
Prompt injection	any	block	Zero-tolerance in strict templates; fail-closed if the guardrail is unreachable.
Toxicity	low	allow	Forwarded unchanged; detection still logged for trend analysis.