---
title: Should you write custom AI agent policies or rely on framework mappings?
url: https://opencontroller.lyzr.com/blog/custom-ai-agent-policies-vs-framework-mappings
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# Should you write custom AI agent policies or rely on framework mappings?

**Use both, for different jobs. Built-in mappings to OWASP, MITRE ATLAS, NIST AI RMF and ISO/IEC 42001 give you a baseline and evidence in a language auditors recognise. Custom policies encode what only your organisation knows: your data classes, your approval thresholds, your vendors and your business rules.** Most enterprises start with the mappings and add five to twenty custom rules in the first quarter.

The question comes up as soon as a security team sees a framework dashboard: "Can we add our own policies on top of OWASP, NIST and ISO?" The answer should always be yes. Here is how to decide what goes where. For background on how a control plane enforces policy, see [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## What do framework mappings actually give you?

Frameworks describe risks and outcomes, not your environment.

- **OWASP Top 10 for Agentic Applications** names ten agent risk classes, ASI01 to ASI10, from Agent Goal Hijack to Rogue Agents.
- **MITRE ATLAS** catalogues adversary tactics and techniques against AI systems.
- **NIST AI RMF** organises outcomes into four functions: Govern, Map, Measure, Manage.
- **ISO/IEC 42001** specifies requirements for an AI management system, with Annex A controls.

A mapping links a technical check (for example, "agent has a write tool with no scope limit") to one or more of these items. That is valuable for two reasons: it gives coverage you did not have to design, and it produces evidence that maps to an external reference. It does not know that your payments team forbids agents from issuing refunds over a set amount.

## When are built-in mappings enough, and when do you need custom rules?

| Situation | Built-in mappings | Custom policy |
| --- | --- | --- |
| First 90 days, building a baseline | Enough | Optional |
| Audit or board asks "which framework do we follow?" | Enough, as evidence | Not needed |
| You have internal data classes (e.g. Restricted, Confidential) | Not enough | Needed |
| Sector rules (payments limits, clinical content, trading) | Not enough | Needed |
| Approved vendor and model list | Not enough | Needed |
| Region or residency rules | Partly | Usually needed |
| Agent-specific business limits | Not enough | Needed |

The test is simple: if the rule needs a fact that only exists inside your company, it is custom.

## What are five examples of organisation-specific rules?

1. **Approved model list.** Block any agent call to a model not on the approved register; alert the owner and the AI governance team. Maps to NIST AI RMF Govern and to your third-party risk process.
2. **Restricted data never leaves region.** If a prompt contains data tagged Restricted, route only to models hosted in the approved region, otherwise block. See [how to enforce data residency for LLM traffic](/blog/how-to-enforce-data-residency-for-llm-and-agent-traffic).
3. **No write tools without an owner review.** Any new MCP server or tool with write, delete, send or pay capability stays disabled until the agent owner and security approve it in the registry.
4. **Customer-facing groundedness floor.** Customer-facing agents must pass a groundedness score threshold on sampled production traffic; below it for a sustained window, raise severity and notify the owner.
5. **Leaver rule.** If an agent's named owner is deprovisioned in the identity provider, quarantine the agent until a new owner is assigned. See [what happens to AI agents when the owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves).

## How do you keep custom policies manageable?

- **Write them as code.** Store them in Git, review changes in pull requests, and version them like application code.
- **Tag each custom rule with framework IDs where they fit.** Rule 3 above also supports OWASP ASI02 (Tool Misuse and Exploitation). That way custom rules add to framework evidence instead of sitting outside it.
- **Give every rule a severity and an action.** Alert, block, or quarantine. See [how to set severity levels for AI agent policies](/blog/how-to-set-severity-levels-for-ai-agent-policies).
- **Start in observe mode.** Run a new rule as alert-only for one to two weeks, check false positives, then enforce.
- **Retire rules.** Review the list quarterly. Rules that never fire, or always fire, need rewriting.

## How Opencontroller does this

Opencontroller uses policy as code, extensible, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Policies apply across the agent estate, and enforcement depth depends on the connector and traffic path: in the gateway path it can allow, route, redact or block; through connected platforms it can apply per-agent Block and Quarantine. These mappings produce evidence for frameworks; they do not make you compliant on their own. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does mapping to ISO/IEC 42001 make us certified?**
No. Certification comes from an accredited audit of your AI management system. Mappings provide evidence for that audit.

**How many custom policies should we start with?**
Five to ten covering your data classes, approved models and write tools is a practical start.

**Should custom policies block or alert?**
Start with alert, measure false positives, then move high-confidence rules to block.

## Key facts

Framework mappings (OWASP Top 10 for Agentic Applications ASI01-ASI10, MITRE ATLAS, NIST AI RMF Govern/Map/Measure/Manage, ISO/IEC 42001 Annex A) give a baseline and auditor-recognisable evidence. Custom AI agent policies encode organisation-only facts: data classes, approved models, residency, write-tool approval, groundedness floors and owner-leaver rules. Write custom policies as code, tag them with framework IDs, start in observe mode, then enforce. Opencontroller by Lyzr provides extensible policy as code with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Pricing starts from $250K per deployment per year.

## Related

- [How should you set severity levels for AI agent policies?](https://opencontroller.lyzr.com/blog/how-to-set-severity-levels-for-ai-agent-policies.md)
- [How do you replace AI governance spreadsheets and questionnaires with continuous evidence?](https://opencontroller.lyzr.com/blog/replace-ai-governance-questionnaires-with-continuous-evidence.md)
- [What should a CISO require before an AI agent goes to production?](https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
