Skip to content
sinc-LLMBook a 15-minute call

OWASP Top 10 for LLM Applications 2026: The List, What Changed and Where to Start

The 2026 OWASP Top 10 for LLM Applications, entry by entry: what moved since 2025, one control to start with for each risk and a worked example.

On this page

By Mario Alexandre · November 19, 2026 · 8 min read

The short answer: the current OWASP Top 10 for LLM Applications is the 2026 edition, published in August 2026. Prompt Injection is still first and Sensitive Information Disclosure still second. Excessive Agency climbed from sixth to third, Unbounded Consumption rose from tenth to sixth, Improper Output Handling fell from fifth to tenth, and System Prompt Leakage was broadened and renamed Hidden Context Exposure. For the first time the ranking weighs the community vote (three quarters) against a corpus of real incidents (one quarter). Use the list as a checklist for the parts of your application where the model is a component; when the model calls tools and acts, pair it with the separate OWASP Top 10 for Agentic Applications.

The 2026 list at a glance

The table gives each 2026 entry, its rank in the 2025 list and one control from the entry's own prevention guidance that is a reasonable first step.

2026 Entry 2025 rank One control to start with
LLM01 Prompt Injection 1 Keep credentials and state-changing actions in application code, not in the model
LLM02 Sensitive Information Disclosure 2 Treat tool-call arguments, logs, telemetry and embeddings as outputs under the same redaction rules
LLM03 Excessive Agency 6 Minimize the tools, their functions and their permissions
LLM04 Supply Chain 3 Keep a signed inventory of components (an SBOM) and use models from verifiable sources
LLM05 Data and Model Poisoning 4 Track dataset and model lineage and version every dataset change
LLM06 Unbounded Consumption 10 Rate limits, input size limits and hard spending caps
LLM07 Misinformation 9 Ground claims and check them before any action depends on them
LLM08 Hidden Context Exposure 7 (as System Prompt Leakage) Put no secrets in the system prompt and enforce authorization outside the model
LLM09 Vector and Embedding Weaknesses 8 Permission-aware retrieval and separation of data by trust tier
LLM10 Improper Output Handling 5 Encode output for where it is used and use parameterized queries

How the 2026 ranking was built

Earlier editions were built on practitioner judgment. In the preface to the 2026 edition, the project leads describe a corpus of 7,714 real incidents from public vulnerability databases and an AI-harm database, of which 6,639 carried enough detail to classify. They then compared what practitioners fear with what the incident record shows. The final ranking gives the community vote three quarters of the weight and the incident data one quarter.

The two disagreed in useful places. Prompt Injection is the top risk by vote but falls out of the top 10 when ranked by raw incident count alone. The leads read that as a defense effect: teams fight injection hard, so fewer clean exploits reach a public database. It stays first. Misinformation runs the other way: the vote placed it near the bottom and the incident record near the top, and the blended score seats it seventh.

What changed since 2025

Four moves matter for a security plan:

  • Excessive Agency rose to third, because the vote and the incident record agree that agentic deployments are where damage is landing.
  • Unbounded Consumption rose four places, carried by practitioners who weigh resource and cost exhaustion higher than before.
  • Improper Output Handling fell from fifth to tenth. It did not get smaller in scope: it now also covers the insecure code that assistants generate at scale.
  • System Prompt Leakage became Hidden Context Exposure, which covers any non-user-facing context placed in the model's context window: system prompts, developer instructions, retrieved policy text and tool schemas.

Several entries also absorbed newer risks. Prompt Injection now covers cross-modal attacks that hide instructions in an image or an audio track. Supply Chain now covers a promoted model artifact that is not what it claims to be. Data and Model Poisoning now covers fine-tuning subversion.

The ten entries in brief

LLM01 Prompt Injection. Any input that alters the model's behavior in ways the developer did not intend, whether it comes from the user, retrieved content, tool output, an image, audio, memory or the model's own intermediate reasoning. The entry is direct about the limit: models make no architectural distinction between instructions and data, so there is no clean equivalent to parameterized queries. Its controls assume the injection will sometimes succeed. One of them is the "Rule of Two": an agent that has untrusted input, sensitive data and the ability to change state or communicate externally at the same time needs per-action human approval. Our guide to indirect prompt injection in RAG and agents goes deeper on the retrieval case.

LLM02 Sensitive Information Disclosure. Confidential or regulated data exposed through any channel nobody authorized. The 2026 entry lists more channels than the final answer: tool-call arguments, reasoning traces, retrieved chunks, logs, telemetry, embeddings, and even timing and token length.

LLM03 Excessive Agency. Damaging actions performed in response to unexpected, ambiguous or manipulated model output, whatever made the model misbehave. Its prevention list is concrete: minimize tools, minimize what each tool can do, avoid open-ended tools, minimize permissions, run tools in the user's own context, require user approval, mediate every call, monitor tool use and rate limit it.

LLM04 Supply Chain. Tampered or vulnerable third-party models, datasets, adapters such as LoRA, and conversion or merge pipelines. Controls include a maintained, signed component inventory and red teaming third-party models for your use case before adopting them.

LLM05 Data and Model Poisoning. Manipulated data or model artifacts that embed backdoors or biases at any stage, including fine-tuning, embedding creation and retrieval. The first listed control is tracking dataset and model lineage; data version control is another control in the same list.

LLM06 Unbounded Consumption. Uncontrolled inference that disrupts service, runs up cost or enables model extraction. The entry names cost asymmetry as the defining feature: an attacker can trigger expensive computation at almost no cost. Controls include rate limiting, input size validation and hard spending caps.

LLM07 Misinformation. Incorrect or unsupported output that looks credible enough to drive a human decision, a workflow or an agent action. The controls include a claim, check, act pattern and validating tool calls before they run.

LLM08 Hidden Context Exposure. Extraction or reconstruction of hidden instructions and context. It becomes a security issue when that context contains secrets, policy logic, tool schemas or trust boundaries. The first control is not to put sensitive data in hidden context; the third is to enforce authorization independently of the model.

LLM09 Vector and Embedding Weaknesses. Risks in any system that uses similarity search to decide what the model sees, including RAG, vector-backed memory and semantic caches. The entry offers a compact frame: poisoning makes the system wrong, inversion makes it leak, jamming makes it silent, and access-control failure makes it indiscriminate.

LLM10 Improper Output Handling. Passing model output downstream without validation or encoding. Controls include treating the model like any other user, context-aware output encoding, parameterized queries, a strict Content Security Policy and stripping control characters such as ANSI escape sequences before output reaches a terminal or log.

A worked example: one support assistant, five entries

The example is fictional. A software company runs a support assistant that reads incoming tickets, searches the help center and can issue a refund of up to $50 through a tool. Answers are rendered as Markdown in the agent console.

What the assistant does Entry that applies Control the team adds
Reads ticket text written by anyone LLM01 Ticket text goes in a separate, labeled channel; the refund tool checks the customer's order in code
Has untrusted input, customer data and a refund tool at once LLM01 (Rule of Two) and LLM03 Refunds need one click of approval from an agent until the team has test evidence to relax it
Writes every prompt and response to the application log LLM02 Logs pass through the same redaction as answers
Renders Markdown with images in the console LLM10 The console does not fetch external images from model output
Has a long system prompt with the refund policy and an internal API key LLM08 The key moves to the server; the prompt keeps only policy text the team would accept seeing published

None of these controls needs the model to resist manipulation. Each one limits what happens when it does not.

Testing your application against the list

LLM Security Red-Team tests your LLM application against agreed adversarial scenarios. You receive a threat model, evidence for the tested attacks and a prioritized list of controls to address. The published project price is $42,500, with the threat model and control list in 10 to 14 days; final scope and timing are confirmed before work begins. For the architecture side, see our guide to AI guardrails architecture.

Book a 15-minute call about LLM security testing · Request an estimate by email

Frequently asked questions about the OWASP LLM Top 10

Is there a 2026 version of the OWASP Top 10 for LLM Applications?

Yes. The project's repository lists the 2026 edition as the current release, published August 4, 2026. The 2025 list remains available for reference.

What happened to System Prompt Leakage?

It is now LLM08 Hidden Context Exposure. The new name covers the system prompt and everything else the application places in the model's context without showing it to the user, such as retrieved policy text and tool schemas.

Does the LLM Top 10 cover AI agents?

Partly. The preface draws the line explicitly: the LLM list owns the risk when the model is a component inside your application. Once the model acts, with tools, memory and downstream consequences, the risk moves to the OWASP Top 10 for Agentic Applications, and the two lists are meant to be read together.

Can prompt injection be fixed?

Not fully, according to the entry itself, because models cannot reliably separate instructions from data. The guidance is to design for it: limit what a successful injection can reach, validate output in trusted code and require human confirmation for privileged or irreversible actions.

Sources

The list, rankings, method and controls come from the 2026 final text and preface in the project's public repository and from its publication page; the 2025 ranks come from the 2025 text in the same repository. The support assistant is a fictional example. OWASP Top 10 for LLM Applications 2026 (OWASP GenAI Security Project) · Canonical source repository (GitHub)