Cybernion
← Insights

Security

LLM Security Risks: the OWASP Top 10 Explained

30 Aug 2026 · Updated 30 Aug 2026

The OWASP Top 10 for LLM Applications is the reference list of large language model security risks, published by the OWASP Gen AI Security Project. The 2025 list runs from prompt injection at number one through to unbounded consumption, with system prompt leakage and vector and embedding weaknesses new this edition. It is how an organisation makes AI governance concrete for an LLM application.

Every organisation is building on large language models, and most are doing it without a clear picture of the risks. The OWASP Top 10 for LLM Applications is the reference that fixes that, the LLM equivalent of the web application Top 10.

What is the OWASP Top 10 for LLM Applications?

The OWASP Top 10 for LLM Applications is a consensus list of the most critical security risks in applications built on large language models. It is published by the OWASP Gen AI Security Project, the OWASP effort dedicated to generative AI security, and it is the reference practitioners use to understand and prioritise the threats specific to LLM applications.

It plays the same role for LLM applications that the long standing OWASP Top 10 plays for web applications: a shared, authoritative starting point that developers, security teams and buyers can align on. The project began in 2023 as generative AI adoption accelerated, and the current version is the 2025 list. Importantly, it is a security risk list, not a governance framework, it identifies what can go wrong technically, which is then managed within a broader AI governance approach. The rest of this article walks through the 2025 list, what changed, and how to defend against these risks.

What are the 2025 OWASP LLM Top 10 risks?

The 2025 list covers ten risk categories. Each has an identifier of the form LLM01 through LLM10.

# Risk What it is
LLM01 Prompt injection Crafted input that manipulates the model into unintended behaviour
LLM02 Sensitive information disclosure The model reveals confidential data it should not
LLM03 Supply chain Vulnerabilities in models, data or components the application depends on
LLM04 Data and model poisoning Tampering with training or fine-tuning data to corrupt the model
LLM05 Improper output handling Trusting model output and passing it unchecked to other systems
LLM06 Excessive agency Giving the model too much functionality, permission or autonomy
LLM07 System prompt leakage Exposure of the system prompt and the secrets or logic it contains
LLM08 Vector and embedding weaknesses Flaws in the retrieval and embedding layer, such as poisoned knowledge bases
LLM09 Misinformation The model producing false or fabricated information presented as fact
LLM10 Unbounded consumption Uncontrolled resource use leading to denial of service or runaway cost

Prompt injection stands at number one for the second edition running, because it is both the most common and the hardest to fully prevent. The list deliberately spans the whole application, not just the model: supply chain, output handling and the retrieval layer are all included, because an LLM application is a system, not just a model.

What changed in the 2025 OWASP LLM Top 10?

The 2025 edition reflects how LLM applications have matured, and how attackers have followed. Two risks are new. System prompt leakage was added, recognising that the system prompt often contains sensitive instructions or secrets that should not be exposed and frequently are. Vector and embedding weaknesses was added to cover the retrieval and embedding layer, which has become central as retrieval augmented generation and AI agents have taken off.

The ordering shifted too. Sensitive information disclosure rose to number two, reflecting how often real LLM applications leak data they should not, whether training data, other users’ data, or confidential context. And some categories were reframed, unbounded consumption broadened the earlier model denial of service idea to cover not just availability but uncontrolled resource use and the runaway cost that a poorly bounded LLM application can rack up. The through line of the changes is that the risks now reflect production LLM applications with retrieval, tools and real data, rather than a model in isolation. For anyone who built against the earlier list, the retrieval layer and system prompt handling are the areas most worth revisiting.

How do we defend against LLM security risks?

There is no single control that secures an LLM application, so defence is layered, the same principle as any other system. The core measures map onto the risks above. Treat all input to the model as untrusted and validate it. Segregate and clearly label external or retrieved content so the model does not treat it as trusted instructions. Apply least privilege to what the model and any connected tools can do, which directly limits excessive agency. Require human approval for high impact actions. Filter and validate model output before it is used or passed to another system, which addresses improper output handling.

Beyond the runtime, protect the data used for training and retrieval to counter poisoning, manage the AI supply chain, and guard the system prompt. And because LLM behaviour is probabilistic and hard to fully constrain, adversarial testing and red teaming of the AI system are essential to find weaknesses before attackers do. These defences are not exotic, they are disciplined application of familiar security principles to a new kind of component, which is exactly why the OWASP list frames LLM security as an extension of application security rather than something wholly separate.

How does the OWASP LLM Top 10 fit AI governance and Australian guidance?

The OWASP LLM Top 10 is the technical threat list, and it falls within a broader governance picture. The NIST AI Risk Management Framework and ISO 42001 provide the governance system, how an organisation governs, maps, measures and manages AI risk, while the OWASP list makes the technical threats concrete at the application level.

National guidance reinforces this. The joint international Guidelines for Secure AI System Development, co-sealed by Australia’s ASD, set secure by design expectations across the AI lifecycle, and the ASD’s own AI advice recommends applying it alongside the Essential Eight. So an organisation building an LLM application should use the OWASP LLM Top 10 to understand and test the technical risks, and a governance framework such as ISO 42001 or the NIST AI RMF to manage AI risk at the organisational level. The two work together: governance sets the expectations, and the OWASP list is how they are met in the code. For an organisation standing up AI governance or securing an AI system, Cybernion helps through ISO 42001 and broader advisory.

Frequently asked questions

What is the OWASP Top 10 for LLM Applications?

The OWASP Top 10 for LLM Applications is a consensus list of the most critical security risks in applications built on large language models, published by the OWASP Gen AI Security Project. It is the reference practitioners use to understand and prioritise LLM specific threats, in the same way the original OWASP Top 10 is used for web applications. The current version is the 2025 list.

What are the top LLM security risks in the 2025 list?

The 2025 OWASP LLM Top 10 is prompt injection, sensitive information disclosure, supply chain, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation, and unbounded consumption. Prompt injection remains the number one risk. Sensitive information disclosure rose to second, and system prompt leakage and vector and embedding weaknesses are new entries in the 2025 edition.

What changed in the 2025 OWASP LLM Top 10?

The 2025 edition added two new risks, system prompt leakage and vector and embedding weaknesses, the latter reflecting the rise of retrieval augmented generation. Sensitive information disclosure rose to number two, reflecting how often LLM applications expose data they should not. The framing of some risks was updated, for example unbounded consumption broadened the earlier model denial of service risk to cover uncontrolled resource use and runaway cost.

How do we defend against LLM security risks?

Defences include treating all model input as untrusted and validating it, segregating and labelling external content, applying least privilege to what the model and its tools can do, requiring human approval for high impact actions, filtering and validating model output before it is used, protecting the data used for training and retrieval, and adversarial testing or red teaming of the AI system. No single control is enough, so LLM security uses layered controls like any other system.

How does the OWASP LLM Top 10 fit AI governance?

The OWASP LLM Top 10 is the technical threat list, and it falls within broader AI governance. Frameworks such as the NIST AI Risk Management Framework and ISO 42001 provide the governance system, while national guidance including the joint Guidelines for Secure AI System Development and the ASD’s AI advice set expectations for securing AI. The OWASP list is how an organisation makes those higher level frameworks concrete for an LLM application.


Written by Gaurav Vikash, an ASD endorsed IRAP assessor and senior cyber security leader with 18 years of experience across Australia, the UK and Asia, including CISO and senior security leadership roles. He holds CISSP, CISA, CISM and CRISC and is an ISO 27001 Lead Auditor and ISO 42001 Lead Implementer, and speaks regularly at industry conferences.

Talk to us.

Sources:

  1. OWASP Top 10 for LLM Applications 2025, OWASP Gen AI Security Project, 2025
  2. LLM01:2025 Prompt Injection, OWASP Gen AI Security Project, 2025
  3. AI Risk Management Framework, National Institute of Standards and Technology, 2026
  4. Guidelines for secure AI system development, National Cyber Security Centre, 2023

Last updated: 30 August, 2026

Talk to us about your engagement

A scoped proposal within one business day.