---
title: What PCI DSS evidence do you need when AI agents touch card data?
url: https://opencontroller.lyzr.com/blog/pci-dss-evidence-for-ai-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# What PCI DSS evidence do you need when AI agents touch card data?

**If an agent stores, processes or transmits cardholder data, or can affect the security of systems that do, it is in PCI DSS scope, along with its model endpoint, gateway, memory and logs. The requirements that change most are scoping (12.5.2), stored data protection (Requirement 3), least privilege for system accounts (7.2.5), application account credentials (8.6) and audit logging (Requirement 10). PCI SSC's October 2026 guidance adds non-mandatory advice on human approval and least agency for agents.**

## When is an AI agent in PCI DSS scope?

PCI DSS v4.0.1 scope covers system components that store, process or transmit cardholder data or sensitive authentication data, plus components that connect to them or can affect their security. Using AI does not change that. PCI SSC states plainly that AI use does not exempt an entity from PCI requirements.

An agent pulls these components into scope when card data reaches them:

- The **agent runtime** and its service account
- The **model endpoint**, including a third-party model API, which then becomes a third-party service provider under Requirement 12.8
- The **LLM gateway** or proxy that carries prompts
- **Memory stores, vector databases and caches** holding conversation history
- **Trace and prompt logs**, which often capture full prompts by default

The cheapest control is keeping card data out. Tokenise PANs before they reach the agent, and redact at the gateway so prompts never carry cleartext card numbers. See [how to redact PII in LLM prompts at the gateway](/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway).

## Which PCI DSS requirements change most for agents?

| Requirement | What it asks | Agent evidence |
|---|---|---|
| 12.5.2 Scope confirmed at least annually | Document all locations and flows of account data | Data flow diagram showing every agent, model, gateway, memory store and log that sees card data |
| 3.3.1 / 3.5.1 Stored account data | No SAD after authorisation; PAN unreadable wherever stored | Proof that prompt logs, traces, memory and caches hold no SAD and only masked or tokenised PAN |
| 4.2.1 Data in transit | Strong cryptography over open, public networks | TLS configuration for agent-to-model and agent-to-tool calls |
| 6.5.1 Change management | Changes to in-scope components are approved and tested | Versioned prompts, tools and model changes with approval records |
| 7.2.2 / 7.2.5 Least privilege | Users and system or application accounts get only the access they need | Per-agent permissions, tool allow-lists, periodic review (7.2.5.1) |
| 8.6.1 to 8.6.3 Application accounts | Manage interactive use, no hard-coded passwords, protect and change credentials | Workload identities or vaulted, rotated credentials for each agent |
| 10.2.1 / 10.2.2 Audit logs | Log access to cardholder data with user, event type, time, outcome and origin | Agent action logs with agent ID, initiating user, tool, data touched and result |
| 10.4.1.1 Automated log review | Automated mechanisms review audit logs | Alerts on agent access to card data outside expected patterns |
| 10.5.1 Log retention | Keep at least 12 months, 3 months immediately available | Retention settings for agent and gateway logs |
| 12.10 Incident response | Plan covers relevant incident types | Playbooks for prompt injection, out-of-scope agent actions and agent quarantine |

Requirements 7.2.5, 7.2.5.1, 8.6.x and 10.4.1.1 became mandatory on March 31, 2025, when the v4.0 future-dated requirements took effect.

## What does PCI SSC's AI guidance add?

In October 2026, PCI SSC announced an information supplement, *Security Considerations for AI Systems*. It is explicitly not a set of mandatory requirements, and the PCI standards take precedence where they differ. As reported, it covers:

1. **Two approval models for agent actions.** Case-by-case human approval for each action involving cleartext cardholder data, or monitored autonomy where the organisation defines permitted actions, shutdown triggers and procedures for reversing changes.
2. **Least agency.** Give each agent only the access and capabilities its task needs.
3. **Splitting risky combinations.** Avoid one agent that has sensitive data access, external communication and untrusted input at the same time. Split those across agents with different permissions.
4. **Independent DLP.** Data loss prevention should run outside the AI system so the agent cannot bypass it.
5. **Human accountability.** A suitable person formally accepts responsibility for AI output.

PCI SSC's 2025 AI principles also say AI should not access high-impact secrets such as credentials or cryptographic keys. Read the supplement itself before quoting it in an assessment.

## A PCI checklist for agents

1. Draw the data flow for every agent that could see card data.
2. Tokenise or redact PAN before prompts reach a model.
3. Disable prompt and content recording on in-scope traces, or prove logs are masked.
4. Give each agent its own identity and a reviewed, minimal permission set.
5. Add model providers to your TPSP list with responsibility matrices.
6. Log every agent action with the initiating user, and keep it 12 months.
7. Define permitted actions and a shutdown trigger for each in-scope agent, and test the shutdown.
8. Add prompt injection and rogue-agent scenarios to the incident plan.

See the [least-privilege checklist for AI agents](/blog/least-privilege-checklist-for-ai-agents) and [human approval for high-risk agent actions](/blog/human-approval-for-high-risk-ai-agent-actions).

## How Opencontroller does this

Opencontroller's LLM Gateway sits in the request path to allow, route, redact or block traffic, which is where card data can be stopped before it reaches a model. Delegated on-behalf-of identity ties each agent action to a user. Posture flags risky permissions, and rogue-agent handling monitors behavioural deviation and can quarantine or kill an agent automatically. Policy as code is extensible, so PCI-specific rules can be written alongside built-in framework mappings. Opencontroller does not ship human-approval hold workflows in the gateway, so case-by-case approval needs to live in your agent or workflow tooling. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

For the general method, read [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## Frequently asked questions

**Is an LLM API provider in PCI DSS scope if prompts contain PAN?**
Yes. A provider that receives cardholder data is a third-party service provider under Requirement 12.8, and you need its compliance evidence and a responsibility matrix. Keeping PAN out of prompts avoids this.

**Is PCI SSC's AI guidance mandatory?**
No. PCI SSC states the October 2026 supplement is not mandatory, and PCI DSS takes precedence where they differ. It is useful evidence of good practice.

**How long must agent logs be kept for PCI DSS?**
Requirement 10.5.1 asks for at least 12 months of audit log history, with the most recent 3 months immediately available.

## Key facts

AI agents that store, process or transmit cardholder data are in PCI DSS v4.0.1 scope, along with model endpoints, gateways, memory and logs; model API providers become TPSPs (12.8). Key requirements: 12.5.2, 3.3.1, 3.5.1, 4.2.1, 6.5.1, 7.2.2, 7.2.5/7.2.5.1, 8.6.1-8.6.3, 10.2.1, 10.2.2, 10.4.1.1, 10.5.1 (12 months, 3 immediately available), 12.10. Future-dated v4.0 requirements took effect March 31, 2025. PCI SSC announced the non-mandatory supplement Security Considerations for AI Systems in October 2026 (human approval or monitored autonomy, least agency, independent DLP). Opencontroller LLM Gateway can allow, route, redact or block; no human-approval hold workflow in the gateway. Pricing starts from $250K per deployment per year.

## Related

- [What does least privilege look like for an AI agent?](https://opencontroller.lyzr.com/blog/least-privilege-checklist-for-ai-agents.md)
- [Which AI agent actions need human approval?](https://opencontroller.lyzr.com/blog/human-approval-for-high-risk-ai-agent-actions.md)
- [How do you redact PII from LLM prompts at the gateway?](https://opencontroller.lyzr.com/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
