AI agent security protects your systems from agents that are tricked, over-permissioned, or simply wrong. OWASP’s Top 10 for Agentic Applications, published in December 2025, names the main risks, from goal hijacking and tool misuse to identity abuse and rogue agents. Most of them are answered by five controls: a scoped identity for each agent, approval on consequential actions, treating everything an agent reads as untrusted, an inventory of agents and the tools they reach, and a log of every action with a way to stop it.
The full picture. AI and MCP governance, end to end covers every stage, the systems it runs on and how to build each one.
A chatbot that gives a wrong answer produces a bad paragraph. An agent that gets the same thing wrong can issue a refund, delete a record, or email a customer list to the wrong place, using credentials someone gave it to do its job. That difference is why agent security is its own discipline.
How agent security differs from LLM security
LLM security treats the model as something that takes input and produces text: the concerns are prompt injection, data leakage, and harmful output. Agent security starts where the model gets credentials, tools, memory, and the autonomy to chain steps together. The question shifts from what the model might say to what the agent might do, in which systems, and on whose authority.
The ten risks OWASP names
The OWASP GenAI Security Project built its Top 10 for Agentic Applications with more than 100 contributors, from incidents seen in real systems. The descriptions and controls below are a plain-language summary; the OWASP document has the full detail.
| OWASP risk | What goes wrong | The control that answers it |
|---|---|---|
| ASI01 Agent goal hijack | Content the agent reads, such as an email or web page, redirects what it is trying to do | Treat retrieved content as untrusted, and require approval for consequential actions |
| ASI02 Tool misuse | The agent uses a legitimate tool in a harmful way | Scope each tool to the job, and validate its parameters |
| ASI03 Identity and privilege abuse | The agent reuses, escalates, or leaks credentials it inherited | Give each agent its own identity with short-lived, narrow credentials |
| ASI04 Agentic supply chain vulnerabilities | A compromised tool, plugin, MCP server, or other agent changes behavior | Inventory and approve the MCP servers and tools agents can reach |
| ASI05 Unexpected code execution | Natural-language instructions turn into code that runs where it should not | Sandbox code execution, and keep shell access off by default |
| ASI06 Memory and context poisoning | Bad information enters an agent's memory and shapes later runs | Validate what is written to memory, and let it expire |
| ASI07 Insecure inter-agent communication | Messages between agents are spoofed or tampered with | Authenticate agent-to-agent messages |
| ASI08 Cascading failures | One wrong output spreads through automated pipelines | Rate limits, circuit breakers, and caps on what one run can change |
| ASI09 Human-agent trust exploitation | A confident explanation persuades a person to approve a harmful action | Approval requests that show exactly what will happen, not the agent's summary |
| ASI10 Rogue agents | An agent drifts from its intended behavior or acts on its own | Monitor behavior across runs, and keep a way to stop an agent |
Five controls that cover most of the list
- A scoped identity for every agent. An agent connected with an admin’s credentials can do anything that admin can. Give each agent its own identity, or have it act as the named user who asked, with access limited to what its job needs. See scoping an agent’s tool access to its job and how dynamic authentication works for MCP.
- Approval on consequential actions. Payments, bulk emails, access changes, and deletions should wait for a person who sees exactly what will happen. Unanswered requests should expire to no action. See an approval step that expires to no action and where people belong in the loop.
- Everything an agent reads is untrusted. Email, documents, web pages, and tool results can all carry instructions. Keep what an agent can do independent of what it reads, so a hijacked goal cannot reach a tool the agent was never given.
- An inventory of agents and the tools they reach. You cannot secure agents and MCP servers nobody knows exist. See the shadow AI detection tools compared, and the security risks MCP brings to the enterprise.
- A log of every action, and a way to stop. Record who triggered each run, which tools it called, and what changed, and keep the ability to switch an agent off. See an MCP audit trail and the agent observability tools compared.
Where Tray fits
Tray covers controls 1, 2, and 5 for agents that run on Tray or call tools through it. Calls through Agent Gateway are tied to a named user, can run on that user’s own credentials, and are scoped to the job each tool does. Sensitive agent actions can require approval. Every agent action, MCP tool call, and workflow run is logged with who triggered it, what ran, and what came back, and the logs stream to Splunk, Datadog, or another SIEM. One role-based access model covers workflows, agents, and MCP tools.
Tray does not detect prompt injection, sandbox code execution, or red-team models. Those sit with runtime AI security tools. For gateways that govern tool access in front of MCP servers, see MCP gateways compared, and for the wider set of governance tools, AI governance tools compared.
Sources
- OWASP GenAI Security Project: Top 10 for Agentic Applications for 2026 (9 December 2025), announcement
- Tray: Agent Gateway, governance and trust
Read on 17 November 2026.