---
title: How does the OWASP Top 10 for LLM Applications differ from the Top 10 for Agentic Applications?
url: https://opencontroller.lyzr.com/blog/owasp-llm-top-10-vs-agentic-top-10
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# How does the OWASP Top 10 for LLM Applications differ from the Top 10 for Agentic Applications?

**The OWASP Top 10 for LLM Applications (2025 edition) lists risks in applications built on a language model, such as prompt injection and sensitive data disclosure. The OWASP Top 10 for Agentic Applications, first published in December 2025, lists risks that appear once a model can plan, use tools, keep memory and act with its own identity. If you run agents, use both: the LLM list still applies to every model call, and the agentic list covers what the agent does with the result.**

Both come from the [OWASP GenAI Security Project](https://genai.owasp.org/); the LLM list is published at [genai.owasp.org/llm-top-10](https://genai.owasp.org/llm-top-10/). Neither is a certification. They are shared vocabularies that security reviewers use to ask what you do about each risk. For a control-by-control evidence table on the agentic list, see [OWASP Top 10 for Agentic Applications evidence](/blog/owasp-top-10-agentic-applications-evidence).

## What is on each list?

| OWASP LLM Top 10 (2025) | OWASP Agentic Top 10 |
|---|---|
| LLM01 Prompt Injection | ASI01 Agent Goal Hijack |
| LLM02 Sensitive Information Disclosure | ASI02 Tool Misuse and Exploitation |
| LLM03 Supply Chain | ASI03 Identity and Privilege Abuse |
| LLM04 Data and Model Poisoning | ASI04 Agentic Supply Chain Vulnerabilities |
| LLM05 Improper Output Handling | ASI05 Unexpected Code Execution |
| LLM06 Excessive Agency | ASI06 Memory and Context Poisoning |
| LLM07 System Prompt Leakage | ASI07 Insecure Inter-Agent Communication |
| LLM08 Vector and Embedding Weaknesses | ASI08 Cascading Failures |
| LLM09 Misinformation | ASI09 Human-Agent Trust Exploitation |
| LLM10 Unbounded Consumption | ASI10 Rogue Agents |

The rows are not paired by number. The mapping below shows how they relate.

## Where do the two lists overlap?

Several agentic risks are the agent-era form of an LLM risk. The difference is the blast radius: in an LLM app, a successful attack mostly changes text; in an agent, it changes actions.

| LLM risk | Related agentic risk | What changes with agents |
|---|---|---|
| LLM01 Prompt Injection | ASI01 Agent Goal Hijack | Injected text redirects a multi-step plan, not one answer |
| LLM06 Excessive Agency | ASI02 Tool Misuse; ASI03 Identity and Privilege Abuse | Split into misuse of tools and misuse of credentials |
| LLM03 Supply Chain | ASI04 Agentic Supply Chain | Adds MCP servers, tools, skills and other agents |
| LLM05 Improper Output Handling | ASI05 Unexpected Code Execution | Output is executed by the agent itself |
| LLM04 Data and Model Poisoning; LLM08 Vector and Embedding Weaknesses | ASI06 Memory and Context Poisoning | Poison persists in agent memory across sessions |
| LLM09 Misinformation | ASI09 Human-Agent Trust Exploitation | People approve harmful actions because the agent sounds sure |
| LLM10 Unbounded Consumption | ASI08 Cascading Failures (partly) | Runaway loops spread across agents and tools |

[What is excessive agency in AI agents?](/blog/what-is-excessive-agency-in-ai-agents) explains why LLM06 needed splitting.

## What is genuinely new in the agentic list?

Three risks have no real counterpart in the LLM list:

1. **ASI07 Insecure Inter-Agent Communication.** Agents call other agents. Messages can be spoofed, replayed or passed on without checks.
2. **ASI08 Cascading Failures.** One agent's fault becomes another agent's input, so errors compound across a workflow.
3. **ASI10 Rogue Agents.** An agent drifts from intended behaviour and keeps acting, whether through compromise, misalignment or bad configuration.

These are estate-level risks. You cannot fix them inside one prompt; they need inventory, identity and runtime control across all agents.

## Which LLM risks do agent teams forget?

Two LLM risks have no agentic twin but still apply to every agent:

- **LLM02 Sensitive Information Disclosure.** Agents retrieve more data than chat apps, so they can leak more. See [prevent memory poisoning in AI agents](/blog/prevent-memory-poisoning-in-ai-agents) for related memory controls.
- **LLM07 System Prompt Leakage.** Agent system prompts often contain tool descriptions and business rules. Never put credentials there.

## What evidence covers both lists?

A single control often answers several rows. Build evidence once:

| Control | Covers | Evidence |
|---|---|---|
| Agent inventory and registry with owners | ASI04, ASI10, LLM03 | Registry export with versions, tools, MCP servers |
| Per-agent identity, least privilege | ASI03, LLM06 | Permission listings, risky-permission findings |
| Runtime guardrails on inputs and outputs | LLM01, ASI01, LLM02, LLM09 | Guardrail decision logs by agent and version |
| Tool approval and argument limits | ASI02, ASI05 | Approved tool lists, denied-call logs |
| Memory and retrieval controls | ASI06, LLM04, LLM08 | Memory write logs, retrieval filters |
| Spend and rate limits | LLM10, ASI08 | Budget alerts, rate-limit events |
| Behaviour monitoring with quarantine | ASI10, ASI08 | Deviation alerts, quarantine records |
| Red-team and multi-turn evals | All | Eval reports per release |

## How Opencontroller does this

Opencontroller's policy as code ships with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, so each policy decision is tagged with the risks it addresses. Posture flags risky permissions, multi-turn and red-team evals test releases, live guardrails check groundedness and task adherence, and rogue-agent handling monitors behavioural deviation with automatic quarantine and kill. This is evidence for OWASP risks, not certification. See [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks), the [capability reference](/capability-reference), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should we use the LLM Top 10 or the Agentic Top 10?**
Both, if you run agents. The LLM list covers each model call; the agentic list covers tools, identity, memory and multi-agent behaviour.

**Is the OWASP Agentic Top 10 a certification?**
No. It is a risk list from the OWASP GenAI Security Project. You can show evidence for each risk, but there is no OWASP certificate.

**What risks are new in the agentic list?**
Insecure inter-agent communication (ASI07), cascading failures (ASI08) and rogue agents (ASI10) have no direct counterpart in the LLM list.

## Key facts

OWASP Top 10 for LLM Applications 2025: LLM01 Prompt Injection to LLM10 Unbounded Consumption. OWASP Top 10 for Agentic Applications (first published Dec 2025): ASI01 Agent Goal Hijack to ASI10 Rogue Agents. Overlaps: LLM01~ASI01, LLM06~ASI02/ASI03, LLM03~ASI04, LLM05~ASI05, LLM04/LLM08~ASI06, LLM09~ASI09. New in agentic: ASI07, ASI08, ASI10. Neither is a certification. Opencontroller policy as code has built-in OWASP, MITRE ATLAS, NIST and ISO 42001 mappings. Pricing starts from $250K per deployment per year.

## Related

- [What evidence maps to the OWASP Top 10 for Agentic Applications?](https://opencontroller.lyzr.com/blog/owasp-top-10-agentic-applications-evidence.md)
- [What is excessive agency in AI agents?](https://opencontroller.lyzr.com/blog/what-is-excessive-agency-in-ai-agents.md)
- [How do you stop memory poisoning in AI agents?](https://opencontroller.lyzr.com/blog/prevent-memory-poisoning-in-ai-agents.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
