---
title: What should a CISO require before an AI agent goes to production?
url: https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# What should a CISO require before an AI agent goes to production?

**A CISO should require a go-live gate with twelve criteria, each backed by a specific evidence artefact: a named owner, a registry entry, scoped identity, a reviewed tool and MCP list, a data classification, passed red-team and quality evals, runtime guardrails, cost limits, logging, a kill path and a rollback plan. If an agent cannot produce the artefact, it is not ready.**

Most agent incidents are not exotic attacks. They are agents that shipped with a shared API key, an unreviewed tool, no owner and no way to stop them. A production gate turns "is this agent safe?" into a list of documents someone can check in an hour.

## Why do AI agents need a separate go-live gate?

A normal application release review checks code, dependencies and infrastructure. An agent adds three things that review does not cover:

- **It decides what to do at runtime.** The same prompt can lead to different tool calls on different days.
- **It acts with someone's authority.** It may call APIs as a service account or on behalf of a user.
- **Its behaviour changes without a code change.** A new model version, a new tool or an edited system prompt can change outcomes.

The [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/) lists risks such as goal hijacking, tool misuse and identity and privilege abuse. Each of the criteria below maps to at least one of them.

## What are the twelve go-live criteria?

| # | Criterion | Evidence artefact that satisfies it |
|---|---|---|
| 1 | Named business and technical owner | Registry entry with two owners resolved in your identity provider |
| 2 | Registered with a version | Registry record: agent, version, model, system prompt hash |
| 3 | Scoped identity, no shared keys | Workload identity or on-behalf-of configuration; secrets scan showing no static keys |
| 4 | Reviewed tool and MCP list | Approved tool list with permissions per tool; MCP servers from an allow-list |
| 5 | Data classification | Highest data class the agent reads and writes, signed by the data owner |
| 6 | Red-team results | Multi-turn red-team report covering prompt injection, data exfiltration and tool misuse, with fixes retested |
| 7 | Quality evals | Task success and groundedness scores on a held-out test set, with the pass threshold stated |
| 8 | Runtime guardrails | Policy list in force at go-live (redaction, blocked actions, groundedness checks) |
| 9 | Cost limits | Budget per agent and version, with alert thresholds |
| 10 | Logging and traces | Trace sample showing prompts, tool calls and outcomes reaching your log store |
| 11 | Kill and quarantine path | Tested runbook: who can block the agent, how, and how long it took in the test |
| 12 | Rollback plan | Previous version identified and a tested switch-back |

Keep the gate the same for every agent. Scale the depth of evidence by risk tier instead of skipping rows.

## How do you tier the gate by risk?

Not every agent needs a two-week review. A simple three-tier model works for most enterprises:

1. **Tier 3 – internal, read-only.** Summarises internal documents for staff. Criteria 1–5, 7, 10 and 11 required; red-team can be a standard automated suite.
2. **Tier 2 – internal, takes actions.** Creates tickets, updates records. All twelve criteria; red-team must include tool misuse.
3. **Tier 1 – customer-facing or regulated data.** All twelve, plus a manual red-team, a named approver for high-impact actions and a re-review after any model change.

Write the tier into the registry entry so the gate is reproducible at audit time.

## Who signs off, and what happens after launch?

Sign-off should sit with the owner, security and the data owner. Security should not be the only approver; owners carry the operational risk.

The gate is a snapshot. After launch, three things should keep it honest:

- **Continuous evaluation.** Score a sample of production traffic for groundedness and task adherence, not just pre-release test sets.
- **Drift triggers.** A new model, tool or prompt version should send the agent back through criteria 2, 4, 6 and 7.
- **Posture checks.** Re-scan for risky permissions and configuration on a schedule, because permissions creep.

The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) frames this as ongoing "Measure" and "Manage" work, not a one-time approval.

For the full picture of how registry, identity and runtime controls fit together, see [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane). The [AI agent registration questionnaire](/blog/ai-agent-registration-questionnaire) covers criteria 1–5 in more detail, and [how to red-team AI agents with multi-turn attacks](/blog/how-to-red-team-ai-agents-multi-turn) covers criterion 6.

## How Opencontroller does this

Opencontroller, by Lyzr, produces most of the gate's artefacts from systems you already run. It auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and keeps a governed registry of agents, versions, tools and MCP servers. It supports delegated on-behalf-of identity and agent workload identity federation, runs multi-turn and red-team evals, and flags risky permissions and configuration through Posture. At runtime it applies groundedness and task-adherence guardrails, attributes cost per agent and per version through its LLM Gateway, and can block or quarantine an agent through the connected platform. Policies are written as code with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, which gives you evidence for those frameworks. [Book a demo](https://www.lyzr.ai/book-demo/) to see the gate on your own agents.

## Frequently asked questions

**How long should an AI agent production review take?**
For a read-only internal agent with automated evidence, a day or less. For a customer-facing agent with manual red-teaming, plan for one to two weeks.

**Does every model upgrade need a new review?**
It should re-run registration, tool review, red-team and quality evals. Ownership, data classification and the kill path usually carry over.

**What is the most common reason agents fail the gate?**
Shared or static API keys and missing owners. Both are cheap to fix before launch and expensive to fix after an incident.

## Key facts

A production readiness gate for AI agents has twelve criteria: named owner, versioned registry entry, scoped identity with no shared keys, reviewed tool and MCP list, data classification, red-team results, quality evals, runtime guardrails, cost limits, logging, a tested kill and quarantine path, and a rollback plan. Each criterion should be satisfied by a specific evidence artefact. Tier the depth of evidence by risk: internal read-only, internal action-taking, and customer-facing or regulated. Opencontroller by Lyzr generates much of this evidence through auto-discovery, a governed registry, Posture checks, multi-turn and red-team evals, runtime guardrails, per-agent cost attribution and per-agent Block and Quarantine controls. Pricing starts from $250K per deployment per year.

## Related

- [What should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [How do you red-team AI agents with multi-turn attacks?](https://opencontroller.lyzr.com/blog/how-to-red-team-ai-agents-multi-turn.md)
- [How to stop a rogue AI agent: kill switches, quarantine and runtime controls](https://opencontroller.lyzr.com/blog/how-to-stop-a-rogue-ai-agent.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
