AI agent security frameworks mapped to layers
OWASP, NIST and MITRE publish guidance that applies to AI agents, and each maps onto the five layers. OWASP names the risks, NIST frames governance and adversarial attacks, MITRE ATLAS catalogs attacker techniques, and the MCP specification and OAuth documents govern the tool and identity layers. Read on 16 Nov 2026.
Facts read on .
Which frameworks apply to AI agents?
Five public sources apply: the OWASP Top 10 for Agentic Applications for 2026, the OWASP LLM Top 10, the NIST AI Risk Management Framework with its GenAI profile, MITRE ATLAS, and, for tools, the MCP specification.
OWASP published its agentic list on 9 Dec 2025 after work with more than 100 experts. NIST says the AI RMF 1.0 is being revised under the White House AI Action Plan, so check the current version.
None of these sources is a certification. They are shared vocabularies that auditors, vendors and regulators recognize, so mapping your controls to them saves time when you are asked to show your work. Keep the mapping as a living table rather than a one-time report.
Sources: OWASP Top 10 for Agentic Applications for 2026; OWASP Top 10 for LLM and Gen AI Apps (2025); NIST AI Risk Management Framework; NIST AI 600-1, the GenAI Profile; MITRE ATLAS; MCP specification: security best practices.
How do NIST and OWASP guidance map to the five layers?
OWASP risks map to layers by where the harmful step shows. Excessive agency maps to identity and endpoint controls, supply chain maps to the endpoint and gateway, and exposure of sensitive information maps to the browser and gateway.
NIST's AI RMF is organized around governance and risk management rather than layers, so use it to set the program: owners, risk tolerance and review. Use the layer map to place the controls.
A worked example helps. Take excessive agency: the endpoint control is a pre-tool check, the identity control is a narrow scope and short lifetime, the gateway control is an allow list, and the cloud workspace control is the platform's own connector and data policy. One framework item, four layer controls.
Sources: OWASP LLM06:2025 Excessive Agency; OWASP Top 10 for LLM and Gen AI Apps (2025); NIST AI Risk Management Framework.
What does NIST say about securing agent systems?
NIST's Center for AI Standards and Innovation issued a request for information on securing AI agent systems in January 2026. It says agent systems can plan and take autonomous actions that affect real-world systems.
It names three risk groups: adversarial data such as indirect prompt injection, insecure models such as ones with poisoned data, and harmful actions even without adversarial inputs. Each group maps to a different layer.
Read the request for information as a statement of concerns that NIST wants input on, not as a final standard. It signals which risks the U.S. standards body considers distinct for agent systems, which is useful when you decide what to test and what to ask vendors.
Sources: NIST CAISI request for information on securing AI agent systems.
How does MITRE ATLAS help?
ATLAS gives names to attacker behavior against AI-enabled systems, which helps when you write detections and test plans. On 16 Nov 2026 it listed 16 tactics, 221 techniques and 44 mitigations.
Map each technique you care about to the layer where you would see it, then check that a control exists there.
ATLAS also publishes mitigations and case studies. A case study shows how a real incident moved through techniques, which is a good starting point for a tabletop exercise with your own agents in the role of the target.
Sources: MITRE ATLAS; NIST AI 100-2 E2025, Adversarial Machine Learning.
Which standards govern the identity and tool layers?
OAuth documents govern most of them. RFC 9700 gives current best practice for OAuth 2.0 security, and the MCP specification requires a protected MCP server to act as an OAuth 2.1 resource server.
For cloud platforms, Microsoft documents Conditional Access as its policy engine that combines signals into if-then access decisions.
Authorization details change over time, so cite the version of the specification you used. The MCP specification page read for this guide is the 2025-06-18 revision, and it requires protected MCP servers to publish protected resource metadata.
Sources: RFC 9700, Best Current Practice for OAuth 2.0 Security; MCP specification: authorization; Microsoft Entra Conditional Access overview.
How do you use the mapping in practice?
Build a table with one row per framework item and one column per layer, and mark the control you have in each cell. Empty cells are your gaps.
Ask vendors to mark the same table from their documents. Compare the answers with what the documentation states, not what a sales deck says.
Review the table when a source updates. NIST says the AI RMF 1.0 is being revised, and OWASP and MITRE both update their lists, so set a quarterly reminder to check the sources and note the date read.
Sources: Zero Trust overview, Microsoft Learn.
Sources
- OWASP Top 10 for Agentic Applications for 2026 (genai.owasp.org), read
- OWASP Top 10 for LLM and Gen AI Apps (2025) (genai.owasp.org), read
- NIST AI Risk Management Framework (nist.gov), read
- NIST AI 600-1, the GenAI Profile (nvlpubs.nist.gov), read
- MITRE ATLAS (atlas.mitre.org), read
- MCP specification: security best practices (modelcontextprotocol.io), read
- OWASP LLM06:2025 Excessive Agency (genai.owasp.org), read
- NIST CAISI request for information on securing AI agent systems (nist.gov), read
- NIST AI 100-2 E2025, Adversarial Machine Learning (csrc.nist.gov), read
- RFC 9700, Best Current Practice for OAuth 2.0 Security (rfc-editor.org), read
- MCP specification: authorization (modelcontextprotocol.io), read
- Microsoft Entra Conditional Access overview (learn.microsoft.com), read
- Zero Trust overview, Microsoft Learn (learn.microsoft.com), read