Policy engine
Guardrails as code
Policies compile to the runtime gateway and evaluate inline on every prompt, tool call and completion. Turning a policy off is itself a logged, attributable governance event.
Policies published
12
compiled to gateway 2026.07.3
Currently enforcing
11
1 disabled
Evaluations / 30d
2.61M
mean added latency 54 ms p99
Policy triggers / 30d
2,874
blocks, masks and escalations
POL-001 · All egress
No customer PII outside India
When
- payload matches PAN / Aadhaar / account recognisers
- destination region ≠ ap-south-1
Then
- block request
- raise critical finding
- notify DPO
POL-002 · Servicing agents
No account closure without approval
When
- tool = accounts.close
- actor is agent
Then
- park for human approval
- log to ledger
POL-003 · Credit agents
No loan rejection without explanation
When
- decision = decline
- reason codes missing
Then
- block response
- return adverse-action template
POL-004 · All models
No customer data sent to public LLM
When
- endpoint not in sanctioned registry
Then
- block
- revoke session token
POL-005 · Payments & treasury
High-value transaction requires human approval
When
- amount > ₹10,00,000
- initiated by agent
Then
- escalate to approver
- SLA 30 min
POL-006 · All prompts & logs
PII masking required
When
- recogniser hit on stored payload
Then
- mask in transit
- mask in ledger
POL-007 · All channels
Sensitive prompts blocked
When
- prompt classified restricted-topic
Then
- block
- coach user
POL-008 · Retrieval + tools
Prompt injection protection
When
- injection classifier score > 0.72
Then
- strip untrusted segment
- block on repeat
POL-009 · Customer-facing
Toxicity detection
When
- toxicity > 0.4 on output
Then
- rewrite
- warn agent
POL-010 · All channels
Jailbreak detection
When
- known jailbreak signature
- role-play override attempt
Then
- block
- open security case
POL-011 · AML agents
Narrative grounding required
When
- citation coverage < 80%
Then
- warn
- attach source list
POL-012 · Wealth copilots
Suitability disclosure on advice
When
- output contains product recommendation
Then
- append disclosure
- log consent