AI Agent Security Solutions

AI agent security risks, by layer

The main AI agent security risks are indirect prompt injection, excessive agency, over-broad credentials, risky tools and MCP servers, sensitive data leaving, and runaway consumption. Each sits at a layer where a control is best placed: tool hooks at the endpoint, scopes in identity, allow lists at the gateway and platform settings in the cloud workspace.

Facts read on .

What are the main AI agent security risks?

The main risks are manipulated inputs, excessive agency, misused credentials, an untrusted tool supply chain, data leaving, and unbounded consumption. NIST's request for information on securing AI agent systems groups them as models meeting adversarial data such as indirect prompt injection, use of insecure models, and models taking harmful actions without any attack.

OWASP's 2025 LLM list names prompt injection, exposure of sensitive information, supply chain, excessive agency and unbounded consumption, and its Top 10 for Agentic Applications for 2026 was built with more than 100 experts for agents that plan and act.

Not every risk needs a new product. Some are old problems with new reach: an over-privileged service account is dangerous whether a person or an agent holds it. What changes is speed and scale, because an agent can try many actions in minutes and chain tools a person would not combine.

Sources: NIST CAISI request for information on securing AI agent systems; OWASP Top 10 for LLM and Gen AI Apps (2025); OWASP Top 10 for Agentic Applications for 2026.

Which risks sit at which layer?

Each risk is best stopped where the harmful step becomes visible. A credential read is visible on the endpoint, a form upload in the browser, an over-broad token in identity, a tool call at the gateway and a connector setting in the cloud workspace.

The list below pairs each risk with its natural layer. Many risks need two layers, such as a gateway block backed by a scoped token.

Pick the three risks that match your agents and write them as one-line scenarios with an owner. A coding-agent team might choose credential theft from the workstation, a poisoned MCP server and an unreviewed package install. A SaaS-builder team might choose an over-shared connector, a runaway loop and data sent to a public model.

Sources: OWASP Top 10 for LLM and Gen AI Apps (2025); OWASP LLM06:2025 Excessive Agency.

How does excessive agency turn into harm?

Excessive agency turns a bad output into a real action. OWASP says it enables damaging actions in response to unexpected, ambiguous or manipulated model outputs, whatever caused the malfunction.

Triggers include hallucination, direct or indirect prompt injection, a malicious extension and, in multi-agent systems, a compromised peer agent. The fix is fewer tools, narrower permissions and approval for high-impact actions.

Hallucination belongs on the list because it needs no attacker. An agent that invents a file path or a command argument can delete the wrong thing. OWASP notes that excessive agency can follow from a poorly performing model as well as from an attack, so limits apply even when you trust every input.

Sources: OWASP LLM06:2025 Excessive Agency.

What are the risks from tools and MCP servers?

Tools and MCP servers extend what an agent can do, so a compromised or over-privileged one extends the damage. The MCP specification documents a confused deputy problem in MCP proxy servers that connect to third-party APIs.

Treat each server as a dependency: approve it, authorize it, scope it and log its calls.

Approving a server is not a one-time event. Track the publisher, the version, the permissions it asks for and the data it can return. A server that was safe at approval can change with an update, so re-check on every version bump.

Sources: MCP specification: security best practices; MCP specification: authorization.

Where can you read attacker techniques for AI systems?

MITRE ATLAS is a public knowledge base of adversary tactics and techniques involving AI. On 16 Nov 2026 its home page counted 16 tactics, 221 techniques, 44 mitigations and 77 case studies.

NIST's adversarial machine learning taxonomy, AI 100-2 E2025, adds a shared vocabulary for attack types and mitigations. Use both to name risks in your own register.

Use technique names in tickets and detections so that analysts, vendors and auditors share terms. When a vendor says it detects a technique, ask for the technique identifier and the layer where it sees the behavior.

Sources: MITRE ATLAS; NIST AI 100-2 E2025, Adversarial Machine Learning.

What do the risks mean for buying?

Write each risk as a question for the vendor: at which layer does the control act, and does it stop the action or only alert? Vendors differ widely on both.

The Agent Control Point Map shows the point in an agent's run where each vendor acts, from before start to after the fact.

Ask also what the product records when it cannot decide. A control that quietly allows an unknown action leaves a blind spot, while one that asks or denies by default leaves a work-around that people will want to remove. Pick the default on purpose for each action class.

Sources: Zenity: runtime boundaries.

Sources