Skip to content
All articles
AI security

The OWASP Top 10 for LLM applications, read as a buyer's checklist

Five of the ten entries in the 2025 list depend only on the application, wherever it is hosted. All ten are questions you can put to any supplier, cloud or on-premise.

BastionUpdated 4 min read

Language models introduce failure modes that firewalls and antivirus were never designed to see. OWASP maintains the reference list, and the 2025 edition names ten risks. OWASP describes the scope as “risks, vulnerabilities and mitigations for developing and securing generative AI and large language model applications across the development, deployment and management lifecycle”. That scope names development, deployment and management. It does not name hosting.

What each of the ten risks means for a business

  • LLM01:2025 Prompt Injection. An attacker writes text that the model treats as an order. It has already produced a real vulnerability in a shipped product.
  • LLM02:2025 Sensitive Information Disclosure. The assistant reveals something it should not, from its training, its context or a document it was allowed to retrieve.
  • LLM03:2025 Supply Chain. The model weights, the datasets, the inference engine and every library underneath arrive from somewhere, like any dependency.
  • LLM04:2025 Data and Model Poisoning. Someone contaminates what the system learns from, in training, fine-tuning or retrieval.
  • LLM05:2025 Improper Output Handling. Generated text is treated as safe input for another system. An answer that becomes a database query, a shell command or an email is an untrusted string.
  • LLM06:2025 Excessive Agency. The model has more power than the task needs, and that multiplies every other entry in the list.
  • LLM07:2025 System Prompt Leakage. Teams routinely put rules, and sometimes secrets, in the system prompt and treat it as private. It is not private.
  • LLM08:2025 Vector and Embedding Weaknesses. In the retrieval layer the index is a second copy of the content, and access to it is access to the source.
  • LLM09:2025 Misinformation. The model is confidently wrong, which is an operational risk wherever an answer is acted on without review.
  • LLM10:2025 Unbounded Consumption. Cost and load have no limit. In a metered service the result is a bill, and on a fixed appliance it is a queue.

Which of the ten does hosting change?

OWASP entriesWhat local deployment changesWhy
LLM01, LLM02Narrows the exfiltration pathAn instruction that succeeds still needs somewhere to send the data
LLM03, LLM08Gives the organisation direct controlThe model, the engine and the index all run on its own site
LLM10Changes its shapeA variable bill becomes a fixed capacity
LLM04, LLM05, LLM06, LLM07, LLM09NothingThe application, the permissions and the review process decide them, identically in a rack or in a region
Entry numbers from the OWASP Top 10 for LLM Applications 2025. The split is this article's reading of the list

The minimum for a thirty-person organisation

  • Authenticate every user; no shared accounts on the assistant
  • Enforce document permissions at retrieval, not in the prompt
  • Validate any output that another system will act on
  • Keep tool and integration permissions to the narrowest set that works
  • Record which model version is running, and where it came from
  • Protect the index and the logs like the documents they contain
  • Decide, in writing, which answers need a human check before they leave the company

NIST's AI Risk Management Framework sits above this as the lifecycle view. It is “intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems”. OWASP tells you what can go wrong in the application; NIST tells you how to organise the work of preventing it.

With Bastion

What Bastion changes

Bastion is a private AI system delivered as one sealed box that runs inside your building. One monthly fee covers the hardware, the model, the hardened operating system and support, and nothing your team types leaves the building.

Bastion answers for the layers it owns: the model and its version, the serving stack, the operating system and the update path, with no outbound connection required to serve a request.

The five entries that hosting does not touch stay with the deployment.

Questions this article answers

What is the OWASP Top 10 for LLM applications?
The reference list of risks specific to applications built on language models. The 2025 edition names 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.
Which OWASP LLM risks does on-premise deployment reduce?
It narrows the exfiltration path for LLM01 and LLM02, gives direct control over LLM03 and LLM08, and turns LLM10 from a variable bill into fixed capacity. It does nothing for LLM04, LLM05, LLM06, LLM07 and LLM09, which are decided by the application and its permissions.
What is the minimum an organisation should do about these risks?
Authenticate every user, enforce document permissions at retrieval rather than in the prompt, validate any output another system will act on, keep tool permissions narrow, record which model version runs, and protect the index and the logs like the documents they contain.

On the record

  1. 1
  2. 2

    NIST

    AI Risk Management Framework

    read 2026-09-22

  3. 3