---
title: What does the CISA AI Data Security guidance mean for AI agents?
url: https://opencontroller.lyzr.com/blog/cisa-ai-data-security-guidance-and-agents
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# What does the CISA AI Data Security guidance mean for AI agents?

**The joint "AI Data Security" information sheet, released on 22 May 2025 by the NSA's AI Security Center, CISA, the FBI and international partner agencies, sets out best practices for securing data used to train and operate AI systems. For AI agents, "operate" is the key word: retrieval sources, agent memory, tool outputs and logs are all operational data that can be poisoned, leaked or drift.** It is voluntary guidance, not a regulation, but it is a useful checklist for producing evidence that agent data is controlled.

## What does the guidance cover?

The full title is *AI Data Security: Best Practices for Securing Data Used to Train & Operate AI Systems*. It is published as a Cybersecurity Information Sheet. It focuses on three risk areas:

1. **Data supply chain risk:** data from third parties, web sources or vendors that may be tampered with or unlawfully sourced.
2. **Maliciously modified ("poisoned") data:** data deliberately altered to change system behaviour.
3. **Data drift:** gradual change in input data that degrades outputs over time.

It frames controls across the AI lifecycle, from planning and design through data collection, model building and use, to operation and monitoring.

## Why does it matter for agents rather than just models?

Most organisations do not train foundation models. They operate agents on top of them. An agent's data surface includes:

- **Retrieval sources** (document stores, vector indexes, wikis) that feed context.
- **Memory** that persists between sessions.
- **Tool outputs** from APIs, web pages and MCP servers, which flow straight into prompts.
- **Prompts and system instructions,** which are effectively configuration data.
- **Logs and traces,** which often contain sensitive data themselves.

Each of these maps directly onto the guidance's three risks. A poisoned wiki page or a malicious web result is a data supply chain problem for an agent, even if no model was ever trained on it.

## How do the best practices map to agent controls?

The information sheet's general best practices include sourcing reliable data and tracking provenance, verifying integrity in storage and transport, digital signatures for trusted revisions, trusted infrastructure, classification and access control, encryption, secure storage, privacy-preserving techniques, secure deletion, and ongoing risk assessment. Mapped to agents:

| Guidance theme | Agent control | Evidence to keep |
|---|---|---|
| Provenance and source tracking | Register every retrieval source and MCP server; record origin for ingested documents | Inventory of knowledge sources and tools with owners |
| Integrity in storage and transport | Hash or sign indexed documents; TLS to tools and models | Integrity check logs; transport configuration |
| Digital signatures for trusted revisions | Version and sign system prompts, policies and agent configs | Change history with approver and version IDs |
| Classification and access control | Per-agent scopes; delegated user identity for data access | Access policies per agent; access review records |
| Encryption and secure storage | Encrypt vector stores, memory and logs at rest | Key management configuration |
| Privacy-preserving techniques | Redact personal data before prompts and logs | Redaction rules and test results |
| Secure deletion | Expire agent memory and logs on schedule | Retention policy and deletion jobs |
| Ongoing risk assessment | Periodic red-team and drift review | Assessment reports and remediation tickets |

Commentary on the sheet also highlights logging data access and inference activity, immutable logs, role-based access and monitoring for unusual shifts in output distributions. Those are the operational controls most agent programmes already need.

## How do you address the three risks for agents?

**Data supply chain**

1. Allow-list retrieval sources and tool endpoints per agent.
2. Treat web and third-party content as untrusted input. See [stopping prompt injection through tool outputs](/blog/stop-prompt-injection-through-tool-outputs).
3. Review vendor terms for how agent data is used and retained.

**Poisoned data**

1. Restrict who can write to knowledge stores an agent reads.
2. Validate what an agent writes to its own memory. See [preventing memory poisoning in AI agents](/blog/prevent-memory-poisoning-in-ai-agents).
3. Red-team with planted documents and tool responses.

**Data drift**

1. Score production traffic continuously, not just at release.
2. Alert on changes in groundedness, refusal rates or tool-error rates.
3. Re-run evaluation suites when sources change.

## Does following it make you compliant?

There is nothing to be compliant with: it is guidance. It can, however, serve as **evidence for** wider frameworks. Its lifecycle and data controls align with the NIST AI RMF's Map, Measure and Manage functions (see [NIST AI RMF for AI agents](/blog/nist-ai-rmf-for-ai-agents)), and with data and logging controls auditors look for under ISO/IEC 42001. For the overall model of controls across agents, see [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, auto-discovered across AWS, Azure and Google Cloud. Posture detects risky permissions and configuration, delegated on-behalf-of identity scopes data access to the user, live groundedness and task-adherence guardrails and LLM judges watch production traffic for drift, and policy as code ships with mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Who published the AI Data Security guidance?**
The NSA's Artificial Intelligence Security Center, CISA and the FBI, with international partner cybersecurity agencies, on 22 May 2025.

**Is the CISA AI Data Security guidance mandatory?**
No. It is a voluntary Cybersecurity Information Sheet. Organisations use it as a benchmark and as evidence for broader frameworks.

**What are the three data risks it highlights?**
Data supply chain risk, maliciously modified (poisoned) data, and data drift.

## Key facts

The joint Cybersecurity Information Sheet "AI Data Security: Best Practices for Securing Data Used to Train & Operate AI Systems" was released on 22 May 2025 by the NSA AI Security Center, CISA, the FBI and international partners. It is voluntary guidance covering three risks: data supply chain, maliciously modified (poisoned) data and data drift, across the AI lifecycle. For AI agents, operational data such as retrieval sources, memory, tool outputs, prompts and logs falls in scope. Its practices (provenance, integrity, signatures, access control, encryption, privacy techniques, secure deletion, ongoing risk assessment) can serve as evidence for NIST AI RMF and ISO/IEC 42001 programmes.

## Related

- [How do you stop memory poisoning in AI agents?](https://opencontroller.lyzr.com/blog/prevent-memory-poisoning-in-ai-agents.md)
- [How does the NIST AI RMF apply to AI agents?](https://opencontroller.lyzr.com/blog/nist-ai-rmf-for-ai-agents.md)
- [How do you stop prompt injection that arrives through tool outputs?](https://opencontroller.lyzr.com/blog/stop-prompt-injection-through-tool-outputs.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
