Catalog My learning Blog
AI Security Course: OWASP LLM Top 10 2026 and Agentic AI · Phase 1 · Lesson 2 · 4 min read

OWASP Top 10 for LLM Applications 2026 Explained

On this page
  1. The OWASP Top 10 for LLM Applications 2026 list
  2. Three changes in the 2026 list worth understanding
  3. The OWASP Top 10 for Agentic Applications
  4. Turning the OWASP lists into a risk register
  5. Check yourself

The OWASP Top 10 for LLM Applications 2026 list

OWASP published the 2026 edition of the Top 10 for LLM Applications on 4 August 2026. For the first time the ranking mixed two inputs: practitioner votes, weighted 75 percent, and a corpus of 6,639 classified real-world incidents, weighted 25 percent. Eight of the ten entries moved.

2026Risk2025 rankWhat changed
LLM01Prompt Injection1Unchanged, scope widened
LLM02Sensitive Information Disclosure2Unchanged
LLM03Excessive Agency6Up three, because agents now act
LLM04Supply Chain3Down one
LLM05Data and Model Poisoning4Down one, now covers fine-tuning subversion and bad adapters
LLM06Unbounded Consumption10Up four, cost and resource exhaustion
LLM07Misinformation9Up two
LLM08Hidden Context Exposure7Renamed from System Prompt Leakage and broadened
LLM09Vector and Embedding Weaknesses8Down one
LLM10Improper Output Handling5Down five, but scope expanded

Three changes in the 2026 list worth understanding

Excessive Agency jumped to third. In 2025 most deployments were chatbots, and a fooled model produced bad text. Now assistants browse, call tools and act on business systems, so the same mistake moves money or sends email. The risk didn't change, the blast radius did.

Hidden Context Exposure replaced System Prompt Leakage. The old name suggested the problem was the system prompt. The new one covers everything in the model's hidden context: tool schemas, retrieved policy text, workflow criteria and the prompt itself. The lesson is the same: anything in the context can be extracted, so secrets and decision rules don't belong there.

Improper Output Handling fell but grew. It now explicitly includes terminal control characters, the automatic fetching of external resources such as images, and untrusted tool output flowing back through your application. Lower rank, wider scope.

The authors also noted that prompt injection incidents were fewer than practitioners feared, partly because defenses have improved. That's a reason to keep investing, not to relax.

The OWASP Top 10 for Agentic Applications

In December 2025 OWASP published a separate Top 10 for Agentic Applications, for risks that only exist when models act autonomously: agent goal hijack, tool misuse, identity and privilege abuse, agentic supply chain, unexpected code execution, memory and context poisoning, insecure inter-agent communication, cascading failures, human-agent trust exploitation and rogue agents. The 2026 LLM list links to it formally. Phase 3 of this course covers these.

Turning the OWASP lists into a risk register

The lists are prompts for thinking, not a compliance certificate. For each entry ask: are we exposed, what would an attack look like, what controls do we have, and how do we test them? Write the answers down as data in your repository, so they're reviewed like code:

Code
# security/risk_register.py
RISKS = {
    "LLM01 Prompt Injection": {
        "exposure": "high: reads customer messages, KB articles and order notes",
        "controls": ["tool allow-list per intent", "approval for send_email", "spotlighting"],
        "tests": ["promptfoo: indirect-prompt-injection", "garak: promptinject, latentinjection"],
        "owner": "platform-security",
    },
    "LLM03 Excessive Agency": {
        "exposure": "medium: can issue refunds and send email",
        "controls": ["refund limit 50 EUR without approval", "no delete tools"],
        "tests": ["promptfoo: excessive-agency"],
        "owner": "support-eng",
    },
    "LLM08 Hidden Context Exposure": {
        "exposure": "medium: system prompt contains escalation criteria",
        "controls": ["no secrets in prompts", "criteria enforced in code"],
        "tests": ["promptfoo: prompt-extraction"],
        "owner": "support-eng",
    },
    # ... the other seven follow the same shape
}

def gaps():
    return [name for name, r in RISKS.items() if not r["tests"] or not r.get("owner")]

if __name__ == "__main__":
    raise SystemExit(f"Risks without tests or owner: {gaps()}" if gaps() else 0)

Run in CI, it fails when someone adds a risk without a test or an owner. The tests it names are built in Phase 5.

Check yourself

Prompt injection (LLM01), misinformation (LLM07) and improper output handling (LLM10), since the bot can still be steered to give wrong answers or emit harmful links and markup. Excessive agency barely applies without tools.

LLM08 covers all hidden context, including tool schemas, retrieved policy text and decision criteria. If any of those would help an attacker, or a rule is enforced only because the prompt says so, the risk applies.

It's reviewed with code changes, every risk has an owner and named tests, and CI can fail the build when a risk has no test.

It focuses on risks that appear when models take actions: goal hijacking, tool misuse, identity and privilege abuse, memory poisoning and agent-to-agent communication. These matter once an app can change things, not just answer questions.