Learn / AI security fundamentals
Lesson 3 of 510 minReviewed
The OWASP LLM and Agentic Top 10 lists, in practice
Know what each entry of the two OWASP lists means for an engineer, how the lists relate, and how to use the ids in threat models, findings and detections without treating them as a checklist.
The OWASP GenAI Security Project maintains two community lists. The Top 10 for LLM Applications, 2025 edition, covers applications built on large language models [1]. The Top 10 for Agentic Applications for 2026, published in December 2025, covers systems where models plan and act through tools, memory and other agents [2]. Neither is a standard you comply with; both are a shared vocabulary for the failures practitioners keep seeing.
| Entry | What goes wrong |
|---|---|
| LLM01 Prompt Injection | Direct or indirect input changes the model's behaviour against the developer's intent |
| LLM02 Sensitive Information Disclosure | Secrets, personal or proprietary data come back in output |
| LLM03 Supply Chain | Compromised models, datasets, packages or plugins |
| LLM04 Data and Model Poisoning | Manipulated training, fine-tuning or embedding data |
| LLM05 Improper Output Handling | Output reaches a browser, shell or query without validation |
| LLM06 Excessive Agency | More functionality, permission or autonomy than the task needs |
| LLM07 System Prompt Leakage | Secrets or logic placed in prompts get extracted |
| LLM08 Vector and Embedding Weaknesses | RAG leaks across users or ingests poisoned content |
| LLM09 Misinformation | Confident false output, including hallucinated packages |
| LLM10 Unbounded Consumption | Runaway usage, cost or model extraction |
The agentic list exists because autonomy changes the failure modes. When a model chooses tools, keeps memory and delegates to other agents, the question shifts from what it says to what it does and with whose authority [2]. OWASP's earlier Agentic AI Threats and Mitigations guide (February 2025) laid the threat-model groundwork for it [3].
| Entry | What goes wrong |
|---|---|
| ASI01 Agent Goal Hijack | Injected content redirects the agent's objective or plan |
| ASI02 Tool Misuse and Exploitation | Legitimate tools used destructively or to exfiltrate |
| ASI03 Identity and Privilege Abuse | Agents inherit or escalate credentials beyond the task |
| ASI04 Agentic Supply Chain Vulnerabilities | Malicious MCP servers, skills, plugins or prompts |
| ASI05 Unexpected Code Execution | Generated code runs outside the intended sandbox |
| ASI06 Memory and Context Poisoning | Persistent memory seeded to influence later sessions |
| ASI07 Insecure Inter-Agent Communication | Agent messages spoofed or tampered with |
| ASI08 Cascading Failures | One faulty agent propagates through connected workflows |
| ASI09 Human-Agent Trust Exploitation | People over-trust agent output or approve blindly |
| ASI10 Rogue Agents | Agents act outside scope or against their principals |
How the lists relate
Several pairs describe the same root cause at different layers. Prompt injection (LLM01) is the usual mechanism behind goal hijack (ASI01). Excessive agency (LLM06) is what turns a hijack into tool misuse (ASI02) or privilege abuse (ASI03). Supply chain appears in both, but ASI04 adds what agents load at run time: MCP servers, skills and tool descriptions that can change after you approved them. Use the agentic id when the harm comes from an action, and the LLM id when it stays in the model's input or output.
Using the ids well
- Threat models. Tag each row with one or two ids so reviewers can compare systems and spot gaps, such as a design with no LLM05 analysis even though output is rendered as HTML.
- Findings. Put the id in the title of red-team findings next to the concrete impact. The id says which class; the impact says how bad.
- Detections and tests. ThreatDeck's detections and the coverage tracker carry the same ids, so you can see which risks you test and which you only document.
- Not a checklist. A system can be clean on all twenty entries and still fail on its own business logic. The lists say where to look first, not where to stop.
For governance teams, NIST's Generative AI Profile maps the same ground to risk-management actions [4]. For attack sequences and detection mapping, the next lesson covers MITRE ATLAS, which the technique pages on ThreatDeck cross-reference with these ids.
Try it on ThreatDeck
- Technique library Open any of the twenty entries to see recent stories, tools and detections linked to it.
- Coverage tracker Mark which OWASP entries each of your AI systems has been tested for.
- ASI04 Agentic Supply Chain See how MCP and skill compromises map to the agentic list.
Self-check
4 questionsPick an answer to see whether it is right and why. Only your score is saved, in this browser.
-
Prompt injection (LLM01) is the mechanism; misuse of a legitimate tool (ASI02) is the harm. Using both tells readers how it happened and what it did.
-
Once a model plans and acts, the important questions are what it can do and with whose permissions. The agentic list names those failures, such as privilege abuse and memory poisoning.
-
The lists name common failure classes. They help consistency and coverage, but a system can still fail on risks specific to its business logic.
-
Cross-tenant leakage through retrieval is a vector and embedding weakness: access control was not enforced when content was retrieved.
References
- OWASP Top 10 for LLM Applications 2025. OWASP GenAI Security Project, 2024-11.
- OWASP Top 10 for Agentic Applications for 2026. OWASP GenAI Security Project, 2025-12.
- Agentic AI – Threats and Mitigations. OWASP GenAI Security Project, 2025-02.
- NIST AI 600-1: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST, 2024-07.