AI agent security
What is the OWASP agentic top 10? The ten risks OWASP lists for AI agents
OWASP publishes a list of ten security risks for AI agents. This page reports what OWASP wrote, adds a plain reading of each entry, and says what we did not read.
The OWASP agentic top 10 is OWASP's list of ten security risks for AI agents, published as the OWASP Top 10 for Agentic Applications for 2026 (version 2026, December 2025). OWASP describes it as "a globally peer-reviewed framework that identifies the most critical security risks facing autonomous and agentic AI systems." Its ten entries run from ASI01, Agent Goal Hijack, to ASI10, Rogue Agents. Source: OWASP Gen AI Security Project, read 17 November 2026.
plain answer
What the OWASP agentic top 10 is and who wrote it
OWASP's resource page, dated December 9, 2025, calls it the OWASP Top 10 for Agentic Applications for 2026 and says it was "Developed through extensive collaboration with more than 100 industry experts, researchers, and practitioners". The document itself is labelled OWASP Gen AI Security Project, Agentic Security Initiative, Version 2026, December 2025.
The leaders' letter in the document says the list "serves as a compass and go-to reference for security leaders and practitioners to understand and address the Top 10 highest-impact threats and begin their journey". It says each entry has "a brief description, common vulnerabilities, example scenarios, and most importantly, actionable mitigations that teams can start implementing today".
OWASP licenses the document under Creative Commons CC BY-SA 4.0. Its disclaimer says the document "does not, and is not intended to, constitute legal advice".
plain answer
The ten risks, in OWASP's words and in plain terms
Each entry below gives OWASP's own wording first. Lines that start "In plain terms" are ours: a short reading for someone outside security, resting only on the OWASP sentence above it.
ASI01, Agent Goal Hijack: OWASP says "attackers can manipulate an agent's objectives, task selection, or decision pathways". In plain terms: someone gets the agent to work toward their goal instead of the one it was given.
ASI02, Tool Misuse and Exploitation: OWASP says "Agents can misuse legitimate tools due to prompt injection, misalignment, or unsafe delegation or ambiguous instruction". In plain terms: the agent is allowed to use the tool, but uses it in a way nobody intended.
ASI03, Identity and Privilege Abuse: OWASP says this risk "exploits dynamic trust and delegation in agents to escalate access and bypass controls". In plain terms: an agent ends up with more access than its job needs, or borrows access from someone else.
ASI04, Agentic Supply Chain Vulnerabilities: OWASP says these "arise when agents, tools, and related artefacts they work with are provided by third parties and may be malicious, compromised, or tampered with in transit". In plain terms: the parts an agent is built from or plugs into may come from someone you have not vetted.
ASI05, Unexpected Code Execution (RCE): OWASP says "Attackers exploit code-generation features or embedded tool access to escalate actions into remote code execution (RCE), local misuse, or exploitation of internal systems." In plain terms: an agent that can write and run code can be tricked into running someone else's. RCE means remote code execution, where an outsider runs code on your system.
ASI06, Memory & Context Poisoning: OWASP defines context as "any information an agent retains, retrieves, or reuses, such as summaries, embeddings, and RAG stores", and says "adversaries corrupt or seed this context with malicious or misleading data, causing future reasoning, planning, or tool use to become biased, unsafe, or aid exfiltration". In plain terms: bad data gets into what the agent remembers, and it keeps acting on it later. Exfiltration means data being taken out of your systems.
ASI07, Insecure Inter-Agent Communication: OWASP says agents "coordinate via APIs, message buses, and shared memory", and that this risk "occurs when these exchanges lack proper authentication, integrity, or semantic validation". In plain terms: one agent can be fooled by a message that is forged, altered, or that means something other than it appears to.
ASI08, Cascading Failures: OWASP says these occur "when a single fault (hallucination, malicious input, corrupted tool, or poisoned memory) propagates across autonomous agents, compounding into system-wide harm". In plain terms: one mistake in one agent gets passed along and grows.
ASI09, Human-Agent Trust Exploitation: OWASP says agents "can establish strong trust with human users through their natural language fluency, emotional intelligence, and perceived expertise", and "Adversaries or misaligned designs may exploit this trust to influence user decisions, extract sensitive information, or steer outcomes for malicious purposes." In plain terms: a person approves what the agent suggests because it sounds sure of itself.
ASI10, Rogue Agents: OWASP says "Rogue Agents are malicious or compromised AI Agents that deviate from their intended function or authorized scope". In plain terms: an agent that has drifted away from what it was set up to do.
plain answer
How the agentic list relates to OWASP's list for LLM applications
OWASP says the agentic list is built on its other work: "Our focus is on Agentic Apps, but these will not exist in isolation and will be part of developing an LLM App. As a result, our entries reference the OWASP Top 10 for LLM Applications and other relevant standards." The entries draw the line themselves.
Prompt injection. OWASP's LLM01:2025 entry says "A Prompt Injection Vulnerability occurs when user prompts alter the LLM's behavior or output in unintended ways" and separates direct from indirect cases: "Indirect prompt injections occur when an LLM accepts input from external sources, such as websites or files". ASI01 says "Unlike LLM01:2025, which focuses on altering a single model response, ASI01 captures the broader agentic impact". In plain terms: LLM01 is about one answer going wrong, ASI01 is about a whole multi-step job going the wrong way.
Excessive agency. OWASP's ASI02 entry says it "builds on the mitigations of LLM06:2025 (Excessive Agency) by extending them to multi-step agentic workflows and tool orchestration." ASI03 is described as "the agentic evolution of Excessive Agency (LLM06:2025)".
Supply chain and output handling. LLM03:2025 says "LLM supply chains are susceptible to various vulnerabilities, which can affect the integrity of training data, models, and deployment platforms." ASI04 notes that LLM03 focuses on static dependencies, while agentic systems "often compose capabilities at runtime". For code, ASI05 builds on LLM05:2025 Improper Output Handling, which "refers specifically to insufficient validation, sanitization, and handling of the outputs generated by large language models before they are passed downstream to other components and systems".
For the older list, see our plain guides to the OWASP risks for LLM applications and to prompt injection. For what limiting an agent's reach looks like in practice, see our page on AI agent security risk.
plain answer
How OWASP tells the ten entries apart
Several entries sit close together, and the document says where one ends and the next begins. Three examples, each from the entry text.
Tool misuse, identity and code execution. ASI02 covers an agent acting "within its authorized privileges" but using a tool in an unsafe or unintended way. It adds: "If the misuse involves privilege escalation or credential inheritance, it falls under ASI03 (Identity & Privilege Abuse); if the misuse results in arbitrary or injected code execution, it is classified under ASI05 (Unexpected Code Execution)." In plain terms: ask whether the agent used what it was given (ASI02), gained more than it was given (ASI03), or ran code it should not have (ASI05).
Goal hijack, memory poisoning and rogue agents. OWASP says "ASI06 focuses on the persistent corruption of stored context or long-term memory, while ASI10 captures autonomous misalignment that emerges without active attacker control." ASI01 is the case where the attacker changes the goal directly, whether interactively or through documents, templates or external data.
Cascading failures. "ASI08 focuses on the propagation and amplification of faults rather than their origin." The document says to record the initial defect under ASI04, ASI06 or ASI07 and to apply ASI08 when it spreads, and it lists symptoms such as "rapid fan-out where one faulty decision triggers many downstream agents or tasks in a short time".
plain answer
What OWASP says to do about it
The entries carry their own mitigation lists. For ASI01, OWASP's first item says to "Treat all natural-language inputs (e.g., user-provided text, uploaded documents, retrieved content) as untrusted." The second says to minimize the impact of goal hijacking by "enforcing least privilege for agent tools and requiring human approval for high-impact or goal-changing actions".
Other entries name different controls. For ASI03: "Issue short-lived, narrowly scoped tokens per task". For ASI05: "Ban eval in production agents". For ASI06: "Scan all new memory writes and model outputs (rules + AI) for malicious or sensitive content before commit." For ASI10: "Implement rapid mechanisms like kill-switches and credential revocation to instantly disable rogue agents." In plain terms: limit what an agent can touch, check what goes into its memory, and be able to switch it off.
The letter also names "Least-Agency", described as "our advice to organizations to avoid unnecessary autonomy", and says "strong observability becomes non-negotiable: without clear visibility into what agents are doing, why they are doing it, and which tools they are invoking, unnecessary autonomy can quietly expand the attack surface and turn minor issues into system-wide failures." For related governance work, see our pages on agent orchestration platform governance and agentic AI governance.
plain answer
Questions the list raises about your own agents
These questions are ours, not OWASP's. Each is a way to read one entry against an agent inventory, and none is a test or a finding.
- Which agents can act on instructions found in documents, email or web pages (ASI01)?
- Which tools can each agent call, and with what permissions (ASI02)?
- Which agents hold credentials of their own, and which borrow a person's (ASI03)?
- Which tools, plug-ins, agents and protocols come from third parties (ASI04)?
- Can any agent generate and run code (ASI05)?
- What does each agent remember, and who can write to it (ASI06)?
- How do agents talk to each other, and is each message authenticated (ASI07)?
- What happens downstream if one agent is wrong (ASI08)?
- Where does a person approve an action, and do they see enough to disagree (ASI09)?
- How would you notice an agent acting outside its scope (ASI10)?
For a plainer look at the safety side, see our page on AI agent safety.
plain answer
What this page does not cover
This page reports one OWASP document and the OWASP LLM entries it points to. We did not test any product or run any security assessment, and nothing here says that any product, platform or service of ours or anyone else's addresses the list. OWASP does not endorse this page.
We read the OWASP resource page, the version 2026 document, and OWASP's LLM01:2025, LLM03:2025, LLM05:2025 and LLM06:2025 entries on 17 November 2026. We did not read the document's references, its appendices or the exploits and incidents tracker in Appendix D. OWASP's guidance changes, so check its pages for the current version. For how agents differ from fixed scripts, see our page on agentic AI compared with traditional automation.
Sources
Cited sources
- S1: OWASP Top 10 for Agentic Applications for 2026 (OWASP Gen AI Security Project resource page) (December 9, 2025 shown on the page), read 2026-11-18
- S2: OWASP Top 10 For Agentic Applications 2026, Version 2026, December 2025 (PDF, downloaded from the resource page) (Version 2026, December 2025), read 2026-11-18
- S3: LLM01:2025 Prompt Injection (OWASP Gen AI Security Project) (2025 edition; no publication date shown), read 2026-11-18
- S4: LLM06:2025 Excessive Agency (OWASP Gen AI Security Project) (2025 edition; no publication date shown), read 2026-11-18
- S5: LLM03:2025 Supply Chain (OWASP Gen AI Security Project) (2025 edition; no publication date shown), read 2026-11-18
- S6: LLM05:2025 Improper Output Handling (OWASP Gen AI Security Project) (2025 edition; no publication date shown), read 2026-11-18
Straight answers
Frequently asked questions about What is the OWASP agentic top 10? The ten risks OWASP lists for AI agents
What is the OWASP agentic top 10?
It is OWASP's list of ten security risks for AI agents, published as the OWASP Top 10 for Agentic Applications for 2026. OWASP describes it as "a globally peer-reviewed framework that identifies the most critical security risks facing autonomous and agentic AI systems." Source: S1, read 17 November 2026.
Is agent goal hijack the same as prompt injection?
Not quite. OWASP says "Unlike LLM01:2025, which focuses on altering a single model response, ASI01 captures the broader agentic impact". Prompt injection is one way to hijack a goal; the entry also lists deceptive tool outputs and forged agent-to-agent messages. Source: S2 and S3, read 17 November 2026.
How do I tell tool misuse from privilege abuse?
OWASP says ASI02 covers an agent acting within its authorized privileges, and that "If the misuse involves privilege escalation or credential inheritance, it falls under ASI03 (Identity & Privilege Abuse)". Source: S2, read 17 November 2026.
Is the OWASP agentic top 10 legal or compliance advice?
No. OWASP's disclaimer says the document "does not, and is not intended to, constitute legal advice." Source: S2, read 17 November 2026.
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call