OWASP Top 10 for LLM Applications 2025, for people shipping an app
Short answer
The OWASP Top 10 for Large Language Model Applications 2025 is a risk list for apps that call a model, not a scoring system for the model vendor. For an app built with Lovable, Bolt, Cursor, v0, or Replit, the items that show up in production first are LLM02 sensitive information disclosure (API keys in the bundle), LLM07 system prompt leakage (secrets or authorization rules in a prompt), LLM06 excessive agency (an MCP or tool endpoint with no auth), LLM10 unbounded consumption (no rate limit or spend cap), LLM01 prompt injection, and LLM05 improper output handling. The prompt is not an access-control boundary. Enforce that in the database and the route.
Last updated · Sekrd Research
- Source: OWASP Top 10 for LLM Applications, version 2025, published by the OWASP GenAI Security Project.
- LLM07:2025 says the system prompt should not be considered a secret and should not be used as a security control.
- LLM10:2025 Unbounded Consumption expanded the old denial-of-service item to include unexpected cost.
The 2025 list
Names below are the 2025 titles. The PDF’s table of contents drops the year tag on LLM04; the item is Data and Model Poisoning.
- LLM01:2025 Prompt Injection. Untrusted text — a web page you fetched, a ticket, a user’s message — changes what the model does. You cannot reliably instruct the model to ignore it. Constrain what the model is allowed to do, in code, so a successful injection has nothing dangerous to call.
- LLM02:2025 Sensitive Information Disclosure. The app leaks secrets, other users’ data, or the contents of a retrieval index. In AI-built apps the common case is simpler than a jailbreak: the provider key is in the JavaScript. That is the first-hour guide and the env-prefix guide.
- LLM03:2025 Supply Chain. The model, the plugin, the MCP server, or a package the agent added. Pin versions. CVE-2025-55182 is this item in practice for Next.js apps, even though it is not an LLM bug.
- LLM04:2025 Data and Model Poisoning. Training data or documents you retrieve from are attacker-controlled. If your app embeds user content into a knowledge base, do not treat retrieved text as instructions, and do not let one tenant’s documents be retrieved for another.
- LLM05:2025 Improper Output Handling. Model output written into HTML, a shell, or a SQL string is an injection bug in your app. Render model text as text. Do not pass it to
evalor string-build a query with it. - LLM06:2025 Excessive Agency. The model can call tools that change state, read the filesystem, or run SQL, and too many of those tools are reachable. An MCP server that answers
tools/listwith no credentials is this item on the public internet. The checklist is MCP server security. Sekrd’s deep scan reports that condition and does not call the tools. - LLM07:2025 System Prompt Leakage. OWASP’s point is easy to misread. Disclosure of the prompt wording is not the main risk — attackers can infer the rules by using the app. The risk is secrets inside the prompt, and authorization that exists only as a sentence (“you are an admin”, “bypass RLS”). Sekrd flags a client-side system prompt only in those two cases, because flagging every prompt is noise. Move the rule into Row Level Security or into the route handler.
- LLM08:2025 Vector and Embedding Weaknesses. Embeddings that mix tenants, or a vector store with no access check, return one customer’s documents to another. Filter by tenant id in the query, not in the prompt.
- LLM09:2025 Misinformation. The model states something false and your app acts on it or shows it as fact. For a product this is a review and citation problem. Do not let unverified model output trigger a refund, a permission change, or a payment.
- LLM10:2025 Unbounded Consumption. No per-user limit, no max tokens, no hard spend cap. A leaked key or a public chat endpoint becomes an invoice. OpenAI documents that email alerts do not stop requests. Put a hard project limit on, and rate-limit the route that calls the model. Supabase Edge Functions have a documented rate-limit pattern in this guide.
A builder’s order of operations
You do not need a red team to start. Do these before launch, then come back to the rest of the list:
- No provider key in the client bundle.
- Authorization in the database and the handler, not in the prompt and not only in middleware (CVE-2025-29927).
- Tools and MCP servers authenticated, least privilege, no unauthenticated
tools/liston the public origin. - A hard spend cap and a per-user rate limit on every model call.
- Model output treated as data, not as code.
That is also most of the vibe coding security checklist. A free Sekrd scan covers the bundle half (keys, client env names, prompts that contain credentials or an authorization rule, a vulnerable RSC pin). A deep scan adds the unauthenticated MCP probe. Plans are on /pricing.
Frequently asked questions
What is the OWASP LLM Top 10 for 2025?
The OWASP GenAI Security Project’s 2025 list of the ten main risks in applications that use large language models: 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.
Does OWASP LLM07 mean I must hide my system prompt?
No. OWASP says the system prompt should not be treated as a secret or as a security control. Do not put credentials, connection strings, or role and permission rules in it. Enforce those outside the model.
Which OWASP LLM risks matter most for an AI-built app?
The ones that show up in the shipped app: API keys in the frontend (LLM02), authorization written only as a prompt (LLM07), tools or MCP servers reachable without auth (LLM06), and no spend or rate limit on model calls (LLM10). Prompt injection still matters once the model can take actions.
Sources
Don't ship until you're sekrd
Run a free scan to find the vulnerabilities your AI missed.
Scan Your App Free