Skip to content
01

Untrusted input

02

Model context

03

Action validation

04

Tools and data

05

Action record

The list is most useful when each category maps to a concrete system boundary: who can influence the model, what data can enter context, which tools can act, and what control stops a bad path before it reaches production.

Category guides

Ten categories for a technical review.

Each guide explains the risk, walks through an example and lists controls and questions for the engineering team.

LLM01:2025

Prompt Injection

Prompt injection happens when an attacker uses instructions in user input, documents, webpages, tickets, or tool output to change the model's intended behavior.

Open guide

LLM02:2025

Sensitive Information Disclosure

How AI systems can expose secrets, regulated data, customer records, proprietary material, or internal context through outputs, logs, tools, or retrieval.

Open guide

LLM03:2025

Supply Chain

Risks from compromised or unreviewed models, datasets, plugins, packages, MCP servers, prompts, infrastructure, and vendors in an AI application.

Open guide

LLM04:2025

Data and Model Poisoning

Data and model poisoning introduces manipulated data or model artifacts that change the system's behavior. Training, fine-tuning and embedding pipelines all need integrity checks.

Open guide

LLM05:2025

Improper Output Handling

Improper output handling occurs when an application passes generated content to a browser, shell, database, tool or other component without the checks required by that destination.

Open guide

LLM06:2025

Excessive Agency

Excessive agency occurs when an AI system has more autonomy, permissions, tools, or action scope than the task requires.

Open guide

LLM07:2025

System Prompt Leakage

System prompt leakage happens when hidden instructions, policy text, tool rules, internal routing logic, or operational context is exposed through model responses.

Open guide

LLM08:2025

Vector and Embedding Weaknesses

Vector and embedding weaknesses affect RAG systems when embeddings, vector stores, chunking, metadata, retrieval filters, or similarity logic expose or distort context.

Open guide

LLM09:2025

Misinformation

Misinformation occurs when an AI system produces false, unsupported, outdated, or misleading output that users or downstream systems treat as reliable.

Open guide

LLM10:2025

Unbounded Consumption

How attackers or faulty workflows can drive excessive model calls, token use, context growth, tool loops, compute cost, or service disruption.

Open guide

Use it in review

Turn the list into launch criteria.

For each relevant category, write down the affected components, the expected behavior and the test that checks it. Assign an owner and rerun the test when the model or application changes.

Read the official OWASP 2025 PDF

Map each category to runtime components.

Test direct and indirect prompt paths.

Inventory MCP, OAuth, and tool authority.

Review RAG source, tenant, and vector boundaries.

Set launch criteria with named owners.

Keep regression probes after remediation.