Skip to content
← Guides
Guide

What the OWASP Top 10 for Agentic Applications means for enterprises

OWASP’s agentic Top 10 names ten risks from goal hijack to rogue agents. Many come down to a few controls on identity, policy, approval and logs.

By Sycamore.

Key takeaways

  • The OWASP Top 10 for Agentic Applications 2026, published by the OWASP GenAI Security Project in December 2025, names ten security risks specific to AI agents, from Agent Goal Hijack (ASI01) to Rogue Agents (ASI10).
  • OWASP’s opening letter advises least agency. Give an agent no more autonomy than its task needs, because agentic behavior where it is not needed expands the attack surface without adding value.
  • Several of the ten risks begin with content an agent reads, such as documents, tool outputs, memory and messages from other agents. OWASP says agents cannot reliably tell instructions from that content.
  • A few controls recur across many entries. They are a distinct identity with short-lived, task-scoped credentials, policy enforced outside the model, human approval for high-impact actions, and tamper-evident logs.
  • The list is community security guidance, not a law. Use it to threat-model each agent, give every control an owner, and test agents against the risks before and after release.

The OWASP Top 10 for Agentic Applications 2026 is a list of the ten most critical security risks for AI agents, published by the OWASP GenAI Security Project in December 2025.1 It runs from Agent Goal Hijack (ASI01) to Rogue Agents (ASI10), and each entry gives examples, attack scenarios and mitigations.1 For an enterprise, the list is most useful as a checklist for every agent it runs: which of the ten risks apply, which control covers each, and who owns that control.

What is the OWASP Top 10 for Agentic Applications?

It is open security guidance for systems where an AI model plans and acts, not only answers. OWASP describes it as “a globally peer-reviewed framework that identifies the most critical security risks facing autonomous and agentic AI systems”, developed with more than 100 industry experts, researchers and practitioners.1 The release announcement says an expert review board with representatives from bodies including NIST, the European Commission and the Alan Turing Institute evaluated it.2

The document sits on top of OWASP’s earlier work. Its authors say they map it to their Agentic AI Threats and Mitigations taxonomy, which “remains our foundational and detailed taxonomy that this Top 10 relies upon”, and that its entries reference the OWASP Top 10 for LLM Applications, the Top 10 for Non-Human Identities, the CycloneDX standard and the OWASP AI Vulnerability Scoring System.1

One idea runs through the whole list. OWASP extends least privilege to what it calls least agency: “This captures our advice to organizations to avoid unnecessary autonomy; deploying agentic behavior where it is not needed expands the attack surface without adding value.” The same letter says that “strong observability becomes non-negotiable”, because without visibility into what agents do, why and which tools they invoke, unnecessary autonomy can quietly expand the attack surface.1

What are the ten risks?

The ten entries, with OWASP’s IDs and names, are:1

ID Risk What goes wrong
ASI01 Agent Goal Hijack Attackers redirect an agent’s objectives, task selection or decision pathways through prompts, deceptive tool outputs, malicious artifacts, forged agent-to-agent messages or poisoned data.
ASI02 Tool Misuse and Exploitation An agent uses a legitimate tool, within its authorized privileges, in an unsafe or unintended way, such as deleting valuable data, over-invoking costly APIs or exfiltrating data.
ASI03 Identity and Privilege Abuse Attackers exploit delegation chains, inherited or cached credentials and trust between agents to escalate access or act without authorization.
ASI04 Agentic Supply Chain Vulnerabilities Third-party models, tools, plug-ins, datasets, other agents, MCP and A2A interfaces, registries or update channels are malicious, compromised or tampered with in transit, often composed at runtime.
ASI05 Unexpected Code Execution (RCE) Code that an agent generates or is tricked into running leads to host or container compromise, persistence or sandbox escape.
ASI06 Memory & Context Poisoning Attackers corrupt or seed what an agent retains or retrieves, such as summaries, memory, embeddings and RAG stores, so later reasoning, planning or tool use becomes unsafe.
ASI07 Insecure Inter-Agent Communication Messages between agents lack authentication, integrity or semantic validation, so they can be intercepted, spoofed or manipulated.
ASI08 Cascading Failures A single fault, such as a hallucination, malicious input, corrupted tool or poisoned memory, spreads across agents, tools and workflows into system-wide harm.
ASI09 Human-Agent Trust Exploitation People over-trust a fluent, confident agent, and attackers or misaligned designs use that trust to steer decisions, extract information or get a person to carry out the harm.
ASI10 Rogue Agents A malicious or compromised agent deviates from its intended function or authorized scope. Each action may look legitimate while its overall behavior is harmful.

Several entries start in the same place. OWASP explains that “agents and the underlying model cannot reliably distinguish instructions from related content”, which is why a document, a tool result or another agent’s message can redirect an agent.1 ASI09 adds a risk that logs alone can miss: the agent manipulates “the human into performing the final, audited action, making the agent’s role in the compromise invisible to forensics.”1

Which controls address each risk?

Most entries share a small set of controls. The table groups the mitigations OWASP lists under each entry into controls an enterprise can assign to an owner.1 It is a summary; each entry in the document has more specific guidance.

Control Risks it addresses Related guide
Treat inputs, retrieved content, tool outputs and peer-agent messages as untrusted; validate before they act. ASI01, ASI06, ASI07 Govern AI agents in production
Give each agent a distinct identity with short-lived, task-scoped credentials and least-privilege tools. ASI02, ASI03, ASI08 Agent identities and permissions
Check each high-impact action against policy outside the model, before it runs. ASI01, ASI02, ASI03, ASI08 Govern AI agents in production
Require human approval for high-impact, irreversible or privilege-raising actions. ASI01, ASI02, ASI03, ASI05, ASI09 ERP and CRM systems
Inventory AI components, pin versions, verify provenance and allow only approved tools and servers. ASI04 Secure MCP servers
Run generated code in sandboxes with network limits, never as root, and separate generation from execution. ASI05 Prototype to production
Isolate memory by user session and domain, scan memory writes, record provenance and expire unverified entries. ASI06 None yet
Authenticate agents to each other and sign messages, including agent cards used for discovery. ASI07 None yet
Set budgets, rate limits, quotas and circuit breakers that throttle or pause an agent on anomalies. ASI02, ASI08 Control AI agent costs
Keep tamper-evident logs of tool calls, inter-agent messages and policy decisions, and alert on drift. ASI01, ASI02, ASI08, ASI09, ASI10 Govern AI agents in production
Be able to revoke credentials, disable tools or connections and quarantine an agent quickly. ASI04, ASI10 Govern AI agents in production

Three points in the document are worth carrying into design reviews:

  • The model’s output is untrusted too. Under ASI02, OWASP says to “treat LLM or planner outputs as untrusted” and to put a policy enforcement point in front of tool calls that validates intent and arguments, enforces schemas and rate limits, and issues short-lived credentials.1
  • Shared credentials defeat least privilege. Under ASI03: “Without a distinct, governed identity of its own, an agent operates in an attribution gap that makes enforcing true least privilege impossible.”1
  • Plan for stopping agents, not only preventing harm. Under ASI10, OWASP recommends “rapid mechanisms like kill-switches and credential revocation to instantly disable rogue agents”, and quarantine for forensic review. Under ASI04, it recommends a supply chain kill switch that can disable specific tools, prompts or agent connections across all deployments.1

How should an enterprise use the list?

Use it per agent, as part of the review an agent goes through before production and again when it changes:

  1. Know which agents you have. OWASP’s supply chain mitigations include maintaining an inventory of AI components.1 For how to build one, see how to find and inventory every AI agent in your company.
  2. Decide how much autonomy each agent needs. Apply least agency first. For example, decide whether an agent should send messages itself or draft them for a person to review.
  3. Threat-model the agent against the ten entries. Note which apply given its tools, data, memory, other agents and the people who rely on it. For example, ASI05 matters most for agents that generate or run code, ASI07 for agents that work with other agents, and ASI08 for agents whose outputs feed other agents, tools or automated workflows.
  4. Map each applicable risk to a control and an owner. Use the table above as a starting point and record who runs each control.
  5. Test before release and on a schedule. OWASP recommends periodic red-team tests that simulate goal override and verify rollback, under ASI01.1 For building evaluations, see how to take an AI agent from prototype to production.
  6. Reassess when the agent changes. New tools, scopes, models, memory or peer agents change which risks apply.

How does it relate to other OWASP lists?

It builds on them rather than replacing them. OWASP says ASI02 “builds on the mitigations of LLM06:2025 (Excessive Agency) by extending them to multi-step agentic workflows and tool orchestration”, and contrasts ASI01 with LLM01:2025, “which focuses on altering a single model response”.1 Both cross-references point to the 2025 edition of the LLM Top 10.13 OWASP has since published the Top 10 for LLM Applications 2026 on its site, on 3 August 2026, and calls it “the latest community-driven guide to the most critical security risks facing applications powered by large language models”. In that edition, Prompt Injection is LLM01:2026 and Excessive Agency is LLM03:2026.4

For MCP servers, OWASP runs a separate MCP Top 10, which its project page says is at the beta release stage.5 The agentic document also includes appendices that map its entries to the Top 10 for Non-Human Identities and relate them to CycloneDX and AI bills of materials.1

Where Sycamore fits

Sycamore Govern is an AI governance control plane. Its discovery step inventories AI use cases, agents, applications, models, connectors and MCP servers, and external agents join through telemetry, A2A or direct registration. Govern gives every agent an identity, with model, tool and data access routed through policy, and approval gates and scoped controls bound high-impact actions.

Govern monitors live agent requests, model and tool use, and spend, lets teams set budgets and rate limits, and can block or pause behavior outside policy. It can stop an agent by revoking its identity, pausing execution, removing a connection or blocking required model and tool paths, and it keeps a tamper-evident record of approvals, policy checks, access, changes and revocations.

Frequently asked questions

Is the OWASP agentic Top 10 a compliance requirement?

No. It is open security guidance written by the OWASP community and licensed under Creative Commons, and the document says it does not constitute legal advice. It is still a useful checklist when you assess agents, alongside the laws and frameworks that do apply to you. For how those fit together, see the guide to governing AI agents in production.

Does the agentic Top 10 cover MCP?

Yes, as part of the supply chain. ASI04 lists agentic interfaces such as MCP and A2A, and agentic registries, among the third-party components that can be malicious or compromised, and ASI02 notes that tool definitions increasingly come from MCP servers. OWASP also runs a separate MCP Top 10, currently a beta release. For MCP-specific controls, see the guide to securing MCP servers.

Why is ASI01 also called Agent Behavior Hijacking?

OWASP’s December 2025 release announcement referred to “Agent Behavior Hijacking” among the highlighted threats. The published document names ASI01 “Agent Goal Hijack”. Use the document’s IDs and names when you map controls, so your mapping lines up with OWASP’s own mitigations and cross-references.

Where should a company start with the list?

Start with an inventory, because you cannot assess agents you do not know about, and OWASP’s supply chain entry includes keeping an inventory of AI components. Then begin with the agents that can change records, move money or send messages, and with the controls that reduce several risks at once, such as per-agent identity, scoped credentials, policy outside the model and approval for high-impact actions.

Sources

  1. OWASP GenAI Security Project, Agentic Security Initiative, “OWASP Top 10 for Agentic Applications for 2026”, version 2026, December 2025 (pp. 6-7 letter; pp. 9-38 entries ASI01 to ASI10; appendices B and C), https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/

  2. OWASP GenAI Security Project, “OWASP GenAI Security Project Releases Top 10 Risks and Mitigations for Agentic AI Security”, 9 December 2025, https://genai.owasp.org/2025/12/09/owasp-genai-security-project-releases-top-10-risks-and-mitigations-for-agentic-ai-security/

  3. OWASP GenAI Security Project, “LLM06:2025 Excessive Agency”, https://genai.owasp.org/llmrisk/llm062025-excessive-agency/

  4. OWASP GenAI Security Project, “OWASP GenAI LLM Top 10 2026” (August 3, 2026; contents list in the downloadable document), https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/

  5. OWASP, “OWASP MCP Top 10”, https://owasp.org/projects/mcp-top-10

# What the OWASP Top 10 for Agentic Applications means for enterprises

OWASP's agentic Top 10 names ten risks from goal hijack to rogue agents. Many come down to a few controls on identity, policy, approval and logs.

## Key takeaways

- The OWASP Top 10 for Agentic Applications 2026, published by the OWASP GenAI Security Project in December 2025, names ten security risks specific to AI agents, from Agent Goal Hijack (ASI01) to Rogue Agents (ASI10).
- OWASP's opening letter advises least agency. Give an agent no more autonomy than its task needs, because agentic behavior where it is not needed expands the attack surface without adding value.
- Several of the ten risks begin with content an agent reads, such as documents, tool outputs, memory and messages from other agents. OWASP says agents cannot reliably tell instructions from that content.
- A few controls recur across many entries. They are a distinct identity with short-lived, task-scoped credentials, policy enforced outside the model, human approval for high-impact actions, and tamper-evident logs.
- The list is community security guidance, not a law. Use it to threat-model each agent, give every control an owner, and test agents against the risks before and after release.

The OWASP Top 10 for Agentic Applications 2026 is a list of the ten most critical security risks for AI agents, published by the OWASP GenAI Security Project in December 2025.[^owasp-top10] It runs from Agent Goal Hijack (ASI01) to Rogue Agents (ASI10), and each entry gives examples, attack scenarios and mitigations.[^owasp-top10] For an enterprise, the list is most useful as a checklist for every agent it runs: which of the ten risks apply, which control covers each, and who owns that control.

## What is the OWASP Top 10 for Agentic Applications?

It is open security guidance for systems where an AI model plans and acts, not only answers. OWASP describes it as "a globally peer-reviewed framework that identifies the most critical security risks facing autonomous and agentic AI systems", developed with more than 100 industry experts, researchers and practitioners.[^owasp-top10] The release announcement says an expert review board with representatives from bodies including NIST, the European Commission and the Alan Turing Institute evaluated it.[^owasp-release]

The document sits on top of OWASP's earlier work. Its authors say they map it to their Agentic AI Threats and Mitigations taxonomy, which "remains our foundational and detailed taxonomy that this Top 10 relies upon", and that its entries reference the OWASP Top 10 for LLM Applications, the Top 10 for Non-Human Identities, the CycloneDX standard and the OWASP AI Vulnerability Scoring System.[^owasp-top10]

One idea runs through the whole list. OWASP extends least privilege to what it calls least agency: "This captures our advice to organizations to avoid unnecessary autonomy; deploying agentic behavior where it is not needed expands the attack surface without adding value." The same letter says that "strong observability becomes non-negotiable", because without visibility into what agents do, why and which tools they invoke, unnecessary autonomy can quietly expand the attack surface.[^owasp-top10]

## What are the ten risks?

The ten entries, with OWASP's IDs and names, are:[^owasp-top10]

| ID    | Risk                                 | What goes wrong                                                                                                                                                                                       |
| ----- | ------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| ASI01 | Agent Goal Hijack                    | Attackers redirect an agent's objectives, task selection or decision pathways through prompts, deceptive tool outputs, malicious artifacts, forged agent-to-agent messages or poisoned data.          |
| ASI02 | Tool Misuse and Exploitation         | An agent uses a legitimate tool, within its authorized privileges, in an unsafe or unintended way, such as deleting valuable data, over-invoking costly APIs or exfiltrating data.                    |
| ASI03 | Identity and Privilege Abuse         | Attackers exploit delegation chains, inherited or cached credentials and trust between agents to escalate access or act without authorization.                                                        |
| ASI04 | Agentic Supply Chain Vulnerabilities | Third-party models, tools, plug-ins, datasets, other agents, MCP and A2A interfaces, registries or update channels are malicious, compromised or tampered with in transit, often composed at runtime. |
| ASI05 | Unexpected Code Execution (RCE)      | Code that an agent generates or is tricked into running leads to host or container compromise, persistence or sandbox escape.                                                                         |
| ASI06 | Memory & Context Poisoning           | Attackers corrupt or seed what an agent retains or retrieves, such as summaries, memory, embeddings and RAG stores, so later reasoning, planning or tool use becomes unsafe.                          |
| ASI07 | Insecure Inter-Agent Communication   | Messages between agents lack authentication, integrity or semantic validation, so they can be intercepted, spoofed or manipulated.                                                                    |
| ASI08 | Cascading Failures                   | A single fault, such as a hallucination, malicious input, corrupted tool or poisoned memory, spreads across agents, tools and workflows into system-wide harm.                                        |
| ASI09 | Human-Agent Trust Exploitation       | People over-trust a fluent, confident agent, and attackers or misaligned designs use that trust to steer decisions, extract information or get a person to carry out the harm.                        |
| ASI10 | Rogue Agents                         | A malicious or compromised agent deviates from its intended function or authorized scope. Each action may look legitimate while its overall behavior is harmful.                                      |

Several entries start in the same place. OWASP explains that "agents and the underlying model cannot reliably distinguish instructions from related content", which is why a document, a tool result or another agent's message can redirect an agent.[^owasp-top10] ASI09 adds a risk that logs alone can miss: the agent manipulates "the human into performing the final, audited action, making the agent's role in the compromise invisible to forensics."[^owasp-top10]

## Which controls address each risk?

Most entries share a small set of controls. The table groups the mitigations OWASP lists under each entry into controls an enterprise can assign to an owner.[^owasp-top10] It is a summary; each entry in the document has more specific guidance.

| Control                                                                                                         | Risks it addresses                | Related guide                                                                  |
| --------------------------------------------------------------------------------------------------------------- | --------------------------------- | ------------------------------------------------------------------------------ |
| Treat inputs, retrieved content, tool outputs and peer-agent messages as untrusted; validate before they act.   | ASI01, ASI06, ASI07               | [Govern AI agents in production](/guides/how-to-govern-ai-agents/)             |
| Give each agent a distinct identity with short-lived, task-scoped credentials and least-privilege tools.        | ASI02, ASI03, ASI08               | [Agent identities and permissions](/guides/ai-agent-identity-and-permissions/) |
| Check each high-impact action against policy outside the model, before it runs.                                 | ASI01, ASI02, ASI03, ASI08        | [Govern AI agents in production](/guides/how-to-govern-ai-agents/)             |
| Require human approval for high-impact, irreversible or privilege-raising actions.                              | ASI01, ASI02, ASI03, ASI05, ASI09 | [ERP and CRM systems](/guides/ai-agents-and-systems-of-record/)                |
| Inventory AI components, pin versions, verify provenance and allow only approved tools and servers.             | ASI04                             | [Secure MCP servers](/guides/secure-mcp-servers-for-agents/)                   |
| Run generated code in sandboxes with network limits, never as root, and separate generation from execution.     | ASI05                             | [Prototype to production](/guides/ai-agents-prototype-to-production/)          |
| Isolate memory by user session and domain, scan memory writes, record provenance and expire unverified entries. | ASI06                             | None yet                                                                       |
| Authenticate agents to each other and sign messages, including agent cards used for discovery.                  | ASI07                             | None yet                                                                       |
| Set budgets, rate limits, quotas and circuit breakers that throttle or pause an agent on anomalies.             | ASI02, ASI08                      | [Control AI agent costs](/guides/control-ai-agent-costs/)                      |
| Keep tamper-evident logs of tool calls, inter-agent messages and policy decisions, and alert on drift.          | ASI01, ASI02, ASI08, ASI09, ASI10 | [Govern AI agents in production](/guides/how-to-govern-ai-agents/)             |
| Be able to revoke credentials, disable tools or connections and quarantine an agent quickly.                    | ASI04, ASI10                      | [Govern AI agents in production](/guides/how-to-govern-ai-agents/)             |

Three points in the document are worth carrying into design reviews:

- **The model's output is untrusted too.** Under ASI02, OWASP says to "treat LLM or planner outputs as untrusted" and to put a policy enforcement point in front of tool calls that validates intent and arguments, enforces schemas and rate limits, and issues short-lived credentials.[^owasp-top10]
- **Shared credentials defeat least privilege.** Under ASI03: "Without a distinct, governed identity of its own, an agent operates in an attribution gap that makes enforcing true least privilege impossible."[^owasp-top10]
- **Plan for stopping agents, not only preventing harm.** Under ASI10, OWASP recommends "rapid mechanisms like kill-switches and credential revocation to instantly disable rogue agents", and quarantine for forensic review. Under ASI04, it recommends a supply chain kill switch that can disable specific tools, prompts or agent connections across all deployments.[^owasp-top10]

## How should an enterprise use the list?

Use it per agent, as part of the review an agent goes through before production and again when it changes:

1. **Know which agents you have.** OWASP's supply chain mitigations include maintaining an inventory of AI components.[^owasp-top10] For how to build one, see [how to find and inventory every AI agent in your company](/guides/ai-agent-inventory/).
2. **Decide how much autonomy each agent needs.** Apply least agency first. For example, decide whether an agent should send messages itself or draft them for a person to review.
3. **Threat-model the agent against the ten entries.** Note which apply given its tools, data, memory, other agents and the people who rely on it. For example, ASI05 matters most for agents that generate or run code, ASI07 for agents that work with other agents, and ASI08 for agents whose outputs feed other agents, tools or automated workflows.
4. **Map each applicable risk to a control and an owner.** Use the table above as a starting point and record who runs each control.
5. **Test before release and on a schedule.** OWASP recommends periodic red-team tests that simulate goal override and verify rollback, under ASI01.[^owasp-top10] For building evaluations, see [how to take an AI agent from prototype to production](/guides/ai-agents-prototype-to-production/).
6. **Reassess when the agent changes.** New tools, scopes, models, memory or peer agents change which risks apply.

## How does it relate to other OWASP lists?

It builds on them rather than replacing them. OWASP says ASI02 "builds on the mitigations of LLM06:2025 (Excessive Agency) by extending them to multi-step agentic workflows and tool orchestration", and contrasts ASI01 with LLM01:2025, "which focuses on altering a single model response".[^owasp-top10] Both cross-references point to the 2025 edition of the LLM Top 10.[^owasp-top10][^owasp-llm06] OWASP has since published the Top 10 for LLM Applications 2026 on its site, on 3 August 2026, and calls it "the latest community-driven guide to the most critical security risks facing applications powered by large language models". In that edition, Prompt Injection is LLM01:2026 and Excessive Agency is LLM03:2026.[^owasp-llm2026]

For MCP servers, OWASP runs a separate MCP Top 10, which its project page says is at the beta release stage.[^owasp-mcp] The agentic document also includes appendices that map its entries to the Top 10 for Non-Human Identities and relate them to CycloneDX and AI bills of materials.[^owasp-top10]

## Where Sycamore fits

[Sycamore Govern](/govern/) is an AI governance control plane. Its discovery step inventories AI use cases, agents, applications, models, connectors and MCP servers, and external agents join through telemetry, A2A or direct registration. Govern gives every agent an identity, with model, tool and data access routed through policy, and approval gates and scoped controls bound high-impact actions.

Govern monitors live agent requests, model and tool use, and spend, lets teams set budgets and rate limits, and can block or pause behavior outside policy. It can stop an agent by revoking its identity, pausing execution, removing a connection or blocking required model and tool paths, and it keeps a tamper-evident record of approvals, policy checks, access, changes and revocations.

[^owasp-top10]: OWASP GenAI Security Project, Agentic Security Initiative, "OWASP Top 10 for Agentic Applications for 2026", version 2026, December 2025 (pp. 6-7 letter; pp. 9-38 entries ASI01 to ASI10; appendices B and C), https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/

[^owasp-release]: OWASP GenAI Security Project, "OWASP GenAI Security Project Releases Top 10 Risks and Mitigations for Agentic AI Security", 9 December 2025, https://genai.owasp.org/2025/12/09/owasp-genai-security-project-releases-top-10-risks-and-mitigations-for-agentic-ai-security/

[^owasp-llm06]: OWASP GenAI Security Project, "LLM06:2025 Excessive Agency", https://genai.owasp.org/llmrisk/llm062025-excessive-agency/

[^owasp-llm2026]: OWASP GenAI Security Project, "OWASP GenAI LLM Top 10 2026" (August 3, 2026; contents list in the downloadable document), https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/

[^owasp-mcp]: OWASP, "OWASP MCP Top 10", https://owasp.org/projects/mcp-top-10

## Frequently asked questions

### Is the OWASP agentic Top 10 a compliance requirement?

No. It is open security guidance written by the OWASP community and licensed under Creative Commons, and the document says it does not constitute legal advice. It is still a useful checklist when you assess agents, alongside the laws and frameworks that do apply to you. For how those fit together, see the guide to governing AI agents in production.

### Does the agentic Top 10 cover MCP?

Yes, as part of the supply chain. ASI04 lists agentic interfaces such as MCP and A2A, and agentic registries, among the third-party components that can be malicious or compromised, and ASI02 notes that tool definitions increasingly come from MCP servers. OWASP also runs a separate MCP Top 10, currently a beta release. For MCP-specific controls, see the guide to securing MCP servers.

### Why is ASI01 also called Agent Behavior Hijacking?

OWASP's December 2025 release announcement referred to "Agent Behavior Hijacking" among the highlighted threats. The published document names ASI01 "Agent Goal Hijack". Use the document's IDs and names when you map controls, so your mapping lines up with OWASP's own mitigations and cross-references.

### Where should a company start with the list?

Start with an inventory, because you cannot assess agents you do not know about, and OWASP's supply chain entry includes keeping an inventory of AI components. Then begin with the agents that can change records, move money or send messages, and with the controls that reduce several risks at once, such as per-agent identity, scoped credentials, policy outside the model and approval for high-impact actions.

Published November 18, 2026 by Sycamore.