---
title: What evidence does ETSI EN 304 223 expect for securing AI agents?
url: https://opencontroller.lyzr.com/blog/etsi-en-304-223-evidence-for-ai-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# What evidence does ETSI EN 304 223 expect for securing AI agents?

**ETSI EN 304 223 is a European Standard that sets baseline cyber security requirements for AI models and systems. It has 13 principles grouped into five lifecycle phases: secure design, development, deployment, maintenance and end of life. It does not name agents, but every principle applies to them. The evidence buyers need is per agent: threat models, an asset inventory, audit trails, supply-chain records, monitoring and a decommissioning record.**

ETSI announced the standard in January 2026 (version 2.1.1). It builds on ETSI TS 104 223, which grew out of the UK government's Code of Practice for the Cyber Security of AI. Because it is an EN, a European Standard, it carries more weight in Europe than a technical specification. It also gives teams a shared vocabulary with security reviewers.

## Who does ETSI EN 304 223 apply to?

The standard assigns provisions to stakeholder roles, mainly **developers**, **system operators** and **data custodians**. One organisation can hold several roles. A bank that builds an agent on a vendor model is a developer of the agent and the system operator that runs it. The model vendor is a developer upstream.

Scope covers AI systems built on deep neural networks, including generative AI, that are meant for real deployment. Agents built on large language models fall inside that scope.

## How do the 13 principles map to AI agent evidence?

| Phase | Principle | What it means for an agent | Evidence to keep |
|---|---|---|---|
| Design | 1. Raise awareness of AI security threats | Builders and owners know about prompt injection, tool misuse and data leakage | Training records per role |
| Design | 2. Design for security as well as function | Least privilege for tools and data; an allow-list of actions | Design review with permission scope |
| Design | 3. Evaluate threats and manage risks | Threat model per agent | Threat model mapped to MITRE ATLAS or OWASP |
| Design | 4. Enable human responsibility | A named owner; humans can intervene | Owner record; override and stop paths |
| Development | 5. Identify, track and protect assets | Inventory of agents, models, prompts, tools, keys | Live inventory with versions |
| Development | 6. Secure infrastructure | Isolated runtime; secrets out of prompts | Configuration and posture findings |
| Development | 7. Secure the supply chain | Known models, packages and MCP servers | AI-BOM; approved MCP server list |
| Development | 8. Document data, models and prompts | What changed in each version | Version history with prompt and model diffs |
| Development | 9. Conduct testing and evaluation | Security testing before release | Red-team and eval results per version |
| Deployment | 10. Communicate with end users and affected people | Users know they are dealing with an agent and its limits | Usage notes; disclosure text |
| Maintenance | 11. Maintain regular updates and mitigations | Patch models, tools and guardrails | Change log; re-test after updates |
| Maintenance | 12. Monitor system behaviour | Detect abnormal actions and drift | Runtime logs, alerts, deviation reports |
| End of life | 13. Ensure proper data and model disposal | Retire the agent, revoke credentials, delete memory | Decommission record; revoked keys |

Check the published standard for the exact wording and the provisions under each principle. The table summarises intent, not the normative text.

## Which principles are hardest for AI agents?

Three tend to fail first.

1. **Principle 5, assets.** Agents appear in cloud projects, SaaS platforms and on developer laptops. A static spreadsheet goes stale within weeks. See [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).
2. **Principle 7, supply chain.** Agents pull in models, tool packages and MCP servers at runtime. You need a list of what each agent depends on. See [what an AI bill of materials is](/blog/what-is-an-ai-bill-of-materials).
3. **Principle 13, end of life.** Retired agents often keep live API keys and OAuth tokens. Decommissioning must revoke credentials and delete stored memory, not only switch off the endpoint.

## ETSI EN 304 223 checklist for an agent programme

1. Assign developer, operator and data-custodian roles for each agent.
2. Write a threat model per agent class before release.
3. Keep a live inventory of agents, versions, models, tools, MCP servers and keys.
4. Record an AI-BOM for each version.
5. Red-team every material agent before release and after major changes.
6. Log every tool call with the agent identity and the user it acted for.
7. Alert on behavioural deviation, and test that you can quarantine an agent.
8. Retire agents with a checklist: revoke credentials, delete memory, archive logs.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and keeps them in a governed registry of agents, versions, tools and MCP servers (Principles 5, 7 and 8). Posture detects risky permissions and configurations (Principles 2 and 6). Multi-turn and red-team evals cover Principle 9. Rogue-agent handling monitors for behavioural deviation, raises alerts and can quarantine or kill an agent (Principle 12). Opencontroller Edge extends this to coding agents, skills and local MCP servers on developer devices. Built-in policy mappings cover OWASP, MITRE ATLAS, NIST and ISO 42001, and you can add ETSI controls as custom policy. This is **evidence for** the standard, not a certification. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

Read next: [how to map agent controls to frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) and [threat-modelling agents with MITRE ATLAS](/blog/mitre-atlas-threat-model-ai-agents).

## Frequently asked questions

**Is ETSI EN 304 223 mandatory?**
It is a voluntary European Standard. It can still appear in procurement requirements and supervisory expectations, and it gives security teams a common baseline to assess AI systems against.

**Does ETSI EN 304 223 cover AI agents?**
It does not single out agents, but it covers AI systems built on deep neural networks, including generative AI. Agents built on large language models fall inside that scope.

**How many principles does ETSI EN 304 223 have?**
Thirteen, grouped into five lifecycle phases: secure design, secure development, secure deployment, secure maintenance and secure end of life.

## Key facts

ETSI EN 304 223 (Securing Artificial Intelligence: Baseline Cyber Security Requirements for AI Models and Systems), V2.1.1, announced January 2026; builds on ETSI TS 104 223 and the UK Code of Practice for the Cyber Security of AI. 13 principles across 5 lifecycle phases: secure design (1-4), development (5-9), deployment (10), maintenance (11-12), end of life (13). Roles include developers, system operators and data custodians. Scope: deep neural network AI including generative AI. Opencontroller discovery, registry, Posture, red-team evals and rogue-agent handling provide evidence for the standard.

## Related

- [How do you use MITRE ATLAS to threat-model AI agents?](https://opencontroller.lyzr.com/blog/mitre-atlas-threat-model-ai-agents.md)
- [What is an AI bill of materials (AI-BOM)?](https://opencontroller.lyzr.com/blog/what-is-an-ai-bill-of-materials.md)
- [What evidence maps to the OWASP Top 10 for Agentic Applications?](https://opencontroller.lyzr.com/blog/owasp-top-10-agentic-applications-evidence.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
