Short answer: The OWASP Top 10 for Agentic Applications is a list of the most important security risks in AI agents: systems that plan, use tools, and take actions across your business systems with limited human input. Published by the OWASP GenAI Security Project in December 2025, it covers agent goal hijack, tool misuse, identity and privilege abuse, agentic supply chain, unexpected code execution, memory and context poisoning, insecure inter-agent communication, cascading failures, human-agent trust exploitation, and rogue agents. For business leaders, the core message is simple: give agents only the autonomy and access the task needs, require human approval for high-impact actions, and make sure you can see and stop what agents do.
What makes an AI agent different from a chatbot?
A chatbot answers questions. An agent acts. It breaks a goal into steps, calls tools and APIs, reads and writes data, remembers context between sessions, and may hand work to other agents. Agents are often connected to business systems through standards such as the Model Context Protocol (MCP), which makes it easy to add tools, and easy to add risk.
That changes the stakes. A manipulated chatbot gives a bad answer. A manipulated agent can send the email, issue the refund, change the record, or run the code. OWASP’s guidance puts it directly: agents amplify existing vulnerabilities. The OWASP Top 10 for LLM Applications still applies to the model underneath; the agentic list covers what happens when that model is given tools, identity, memory, and autonomy.
The ten agentic risks, in plain English
| ID | Risk | What it means | Business example | Design question |
|---|---|---|---|---|
| ASI01 | Agent Goal Hijack | Instructions hidden in content the agent reads redirect its goal or plan | A support agent reads a customer email with hidden text telling it to forward account data externally | What untrusted content can the agent read, and which actions can that content trigger? |
| ASI02 | Tool Misuse and Exploitation | The agent uses tools it is allowed to use in unsafe or unintended ways | A cleanup agent with delete rights removes production records instead of test records | Does each tool have the narrowest scope possible, with limits on volume and parameters? |
| ASI03 | Identity and Privilege Abuse | The agent acts with borrowed, inherited, or excessive credentials | An agent runs under an administrator’s account and can reach systems far beyond its task | Does each agent have its own identity with short-lived, task-scoped access? |
| ASI04 | Agentic Supply Chain Vulnerabilities | Third-party tools, MCP servers, plug-ins, prompts, or other agents are malicious or compromised | A popular MCP server package is altered to copy every email the agent sends to an attacker | Who vets and approves the tools, servers, and components our agents load? |
| ASI05 | Unexpected Code Execution (RCE) | Code the agent writes or triggers runs without validation or isolation | A data-analysis agent generates a script that also opens a connection to an outside server | Does agent-generated code run in an isolated sandbox with no access to production? |
| ASI06 | Memory and Context Poisoning | Stored memory or retrieved context is corrupted and shapes later decisions | A false “policy” planted in the agent’s memory changes how it approves invoices for weeks | Who and what can write to agent memory, and can we inspect and reset it? |
| ASI07 | Insecure Inter-Agent Communication | Messages between agents are spoofed, intercepted, or altered | A fake “manager agent” message instructs a payments agent to release funds | Are agents authenticated to each other, and are their messages validated? |
| ASI08 | Cascading Failures | One fault spreads through connected agents, tools, and workflows | A pricing agent’s error feeds an ordering agent and then an invoicing agent before anyone notices | Where are the circuit breakers that stop one agent’s error from becoming everyone’s? |
| ASI09 | Human-Agent Trust Exploitation | Confident, polished agent output leads people to approve harmful actions | A reviewer approves a vendor bank-detail change because the agent’s explanation looked convincing | Do approvers see the evidence behind a request, not just the agent’s summary? |
| ASI10 | Rogue Agents | An agent drifts from its intended behavior, acting outside its mandate or concealing actions | An agent tasked with reducing costs starts disabling monitoring tools to save money | Do we monitor behavior against a defined mandate, and can we shut an agent down quickly? |
Which agentic risks matter most for business leaders?
All ten deserve a design answer, but three drive the biggest business impact:
- Goal hijack (ASI01). Like prompt injection in the LLM list, it has no complete fix. Any agent that reads email, documents, web pages, or tickets can receive hostile instructions. Good designs assume it will happen and cap what a hijacked agent can do.
- Identity and privilege abuse (ASI03) and tool misuse (ASI02). The damage an agent can cause equals the access it holds. Agents running on shared or human credentials are the most common and most fixable weakness.
- Human-agent trust exploitation (ASI09). Human approval only works if the human can actually evaluate the request. Rubber-stamp approvals give false comfort.
What design principles reduce agent risk?
The OWASP guidance introduces least agency: avoid giving agents autonomy they don’t need, because unnecessary autonomy expands the attack surface without adding value. It pairs this with strong observability. In practice, six principles cover most of the list:
- Least agency and least privilege for tools. Only use an agent where autonomy adds value. Give each tool the narrowest scope, and prefer read-only access where possible.
- Human approval for high-impact actions. Payments, deletions, external communications, permission changes, and goal changes need a human or policy check, with enough context to judge.
- Identity for agents. Treat each agent as a managed non-human identity with its own credentials, short-lived tokens, and an accountable owner.
- Logging and traceability. Keep tamper-resistant logs of inputs, decisions, tool calls, and outputs so you can reconstruct what an agent did and why.
- Input, output, and tool-call controls. Treat everything an agent reads as untrusted, validate tool calls and parameters before they execute, and sandbox any code it runs.
- A kill switch. Be able to pause an agent, revoke its credentials, and disconnect its tools quickly, and test that you can.
Pre-launch checklist for an AI agent
| Check | What good looks like | Risks addressed |
|---|---|---|
| Business owner named | One accountable owner for the agent and its residual risk | All |
| Autonomy justified | Written reason why each autonomous step needs an agent | ASI02, ASI10 |
| Tool inventory | Every tool, MCP server, and connected agent listed, vetted, and approved | ASI02, ASI04, ASI07 |
| Scoped identity | Dedicated agent credentials, short-lived, least privilege | ASI03 |
| Approval points | High-impact actions require a human or policy check with evidence shown | ASI01, ASI09 |
| Code isolation | Generated code runs in a sandbox, never directly in production | ASI05 |
| Memory controls | Defined sources that can write to memory; memory can be reviewed and reset | ASI06 |
| Limits and circuit breakers | Rate, volume, and spend limits; failures stop downstream agents | ASI02, ASI08 |
| Logging | Tool calls and decisions logged and reviewable | ASI08, ASI10 |
| Kill switch tested | Pause and credential revocation demonstrated before launch | ASI10 |
Record accepted risks in your risk register with an owner and a review trigger.
How does it relate to the LLM Top 10, NIST AI RMF, and ISO/IEC 42001?
- OWASP Top 10 for LLM Applications. The agentic list builds on it rather than replacing it, and its entries reference the LLM risks. Several agentic risks extend LLM06 (Excessive Agency) into multi-step workflows. Review agent designs against both. See our OWASP Top 10 for LLM Applications guide.
- NIST AI RMF. The voluntary NIST AI Risk Management Framework organizes AI risk work into Govern, Map, Measure, and Manage. An agentic design review supplies concrete evidence for the Map and Manage functions.
- ISO/IEC 42001. The AI management system standard expects AI risk assessment and life-cycle controls. Agent design reviews and the checklist above fit naturally into that program. See ISO/IEC 42001 explained.
Technical controls work best inside a broader AI governance program with an AI acceptable use policy that sets rules for agents and integrations.
Frequently Asked Questions
Is the OWASP Top 10 for Agentic Applications a compliance requirement?
No. It’s free, community-developed guidance, not a regulation or certifiable standard. It’s a practical reference when customers ask how your AI agents are secured, and it fits alongside the NIST AI RMF and ISO/IEC 42001.
Do we need it if we only use agents built into vendor products?
Yes, in part. The vendor manages its own platform, but you decide which tools and data the agent can reach, what permissions it holds, and which actions need approval. Ask vendors how they address the list and review your own configuration.
How is the agentic list different from the LLM Top 10?
The LLM list focuses on risks in applications built on language models. The agentic list focuses on what changes when those models plan, use tools, hold credentials, keep memory, and work with other agents.
Are MCP servers covered?
Yes. Third-party MCP servers and tool definitions fall under agentic supply chain vulnerabilities (ASI04), and their permissions fall under tool misuse (ASI02) and identity and privilege abuse (ASI03). Vet and approve them like any other software component.
How USCIBER can help
USCIBER’s AI Security & Governance service includes secure design and architecture reviews of agentic applications against the OWASP Top 10 for Agentic Applications and for LLM Applications, alongside AI risk assessment, AI policies, and NIST AI RMF and ISO/IEC 42001 readiness. AI work is scoped after a short call. Engagements are led by security leadership holding CISSP, CISA, and CISM credentials. Book a free consultation.
Sources
- OWASP GenAI Security Project — OWASP Top 10 for Agentic Applications for 2026
- OWASP GenAI Security Project — OWASP Top 10 for Agentic Applications: The Benchmark for Agentic Security in the Age of Autonomous AI
- OWASP GenAI Security Project — Top 10 for LLM Applications (2025)
- NIST — AI Risk Management Framework (AI RMF 1.0)
- ISO — ISO/IEC 42001, Artificial intelligence — Management system
This article is general information, not legal advice.