Q-Cyber Magazine

AI · 14 May 2026 · 3 min read

Prompt injection and the OWASP Top 10 for LLM applications

OWASP's 2025 Top 10 for LLM applications places prompt injection first, and treats it as a systems problem: untrusted input can change model behaviour, including through indirect content the model is asked to read.

Abstract graphic of a network of connected nodes in Knowlcom blue

A large language model does not enforce a clean boundary between instructions and data. That fact, more than any single exploit technique, is why prompt injection sits at the top of OWASP's 2025 list for LLM applications.

A systems vulnerability, not only a clever prompt

OWASP is careful about what can be claimed. Prompt injection exists in how models process prompts. Retrieval-augmented generation and fine-tuning improve relevance, but research cited by OWASP shows they do not fully close the vulnerability. It is also unclear, given the way models work, whether there is a fool-proof method of prevention. The practical task is therefore mitigation: constrain behaviour, validate outputs, filter inputs and outputs, enforce least privilege, require human approval for high-risk actions, segregate external content, and test adversarially.

The impact depends on context and on how much agency the application has been given. OWASP's list of possible outcomes includes disclosure of sensitive information or system prompts, manipulated or biased outputs, unauthorised use of functions available to the model, execution of commands in connected systems, and interference with important decisions. Multimodal systems add a further path, because instructions can be hidden in one modality, such as an image, while the accompanying text looks ordinary.

None of this is a how-to. It is a reason to treat the model as an untrusted component inside a larger application, especially when that application can read mail, call tools, write to databases or speak to customers.

Exhibit 1

OWASP Top 10 for LLM Applications 2025

IdentifierRisk
LLM01:2025Prompt Injection
LLM02:2025Sensitive Information Disclosure
LLM03:2025Supply Chain
LLM04:2025Data and Model Poisoning
LLM05:2025Improper Output Handling
LLM06:2025Excessive Agency
LLM07:2025System Prompt Leakage
LLM08:2025Vector and Embedding Weaknesses
LLM09:2025Misinformation
LLM10:2025Unbounded Consumption

Source: OWASP Gen AI Security Project, Top 10 Risk and Mitigations for LLMs and Gen AI Apps (2025), https://genai.owasp.org/llm-top-10/.

What to change in design

The most important companion risk on the 2025 list is often LLM06, excessive agency: a model that can take actions will turn a successful injection into a business event. Least privilege, tool isolation, and human approval for high-risk actions are therefore not optional decorations. They are the difference between a bad paragraph and an unauthorised payment, mail or record change.

Indirect injection should be assumed wherever the application summarises web pages, documents or tickets. External content should be segregated and labelled so that it cannot silently rewrite the system prompt. Outputs that will be executed, rendered or stored should be validated as data, not trusted as instructions.

Implications

  1. Treat the model as untrusted, especially if it can call tools or read untrusted content.
  2. Reduce agency until a human approves actions that move money, change records or send messages.
  3. Segregate retrieved content from system instructions.
  4. Test regularly, including multimodal and indirect cases, and accept that mitigation is ongoing.

What this means for AI

An organisation that has connected a language model to internal tools has created a new class of application security work. OWASP's 2025 list is a sufficient starting agenda. Knowlcom's work on sovereign or on-premises AI can help place that agenda inside an estate the organisation already controls, but it does not remove the need for the controls themselves.

Sources

  1. OWASP Top 10 for LLM Applications 2025 index
  2. LLM01:2025 Prompt Injection