OWASP 2025: Prompt Injection, RAG Documents, and Agent Permissions
How can an external page or document misdirect an AI agent? Apply the OWASP 2025 prompt-injection risk model to tools, permissions, and approvals.
Instructions and data are not the same
An AI application may read a webpage, PDF, or database record to answer a user's question. That material is necessary task data, but not every sentence inside it is a trustworthy instruction. OWASP placed prompt injection first among its 2025 risks for large language model applications because external input can redirect model behavior in unintended ways. A user may supply an unsafe request directly, or an attacker may hide instructions inside apparently ordinary third-party content. Clear system instructions help, but they are not a complete security boundary. A safer architecture does not depend on hoping the model will reject every malicious sentence. It separates untrusted source material from the authority to take actions through enforceable application controls.
Indirect attacks: why a document can be dangerous
Imagine an agent that reads a hotel website to extract facility characteristics. A hidden paragraph on the page could instruct the agent to stop summarizing, access unrelated administration records, or send secrets to an external address. To the publisher it is just text; if the model treats it as an authorized task instruction, a trust boundary has been crossed. Retrieving the correct page through RAG does not solve the problem because the retrieved document may itself be untrusted. Preserve source attribution and document time, treat external text as informational input only, and bind available actions to explicit application-side permissions. Providing context is fundamentally different from granting operational authority to whatever instructions happen to appear in that context.
Give model tools the least possible privilege
Being able to query a database does not imply that an AI assistant should be able to delete arbitrary tables. Design tools around separate security boundaries: search should reach only approved sources, reporting should be read-only, and publishing should require a distinct API with its own validation. Do not transfer the full authority of a user session into unrestricted model actions. Secrets belong on trusted server infrastructure, not inside model-generated text or client-side source files. Every tool call should validate the requested operation, target record, permissions, origin, and current revision. Even when the model produces an incorrect recommendation, deterministic database revision guards and uniqueness constraints must still apply. Least privilege reduces the blast radius of compromised context.
When does human approval matter?
An agent may safely draft a headline, flag a typo, or compare public sources through read-only tools. Deleting records, changing permissions, spending money, sending confidential data, or overwriting large datasets has a very different impact. A model merely claiming in text that approval was granted is not sufficient. The application should provide an independent approval mechanism displaying the concrete target and expected side effects. Authorization should be bound to a specific operation and record identity, not reused for unrelated requests later. Distinguishing routine reversible work from high-impact mutations preserves throughput without sacrificing data integrity. Human review is meaningful only when it controls the actual execution boundary rather than a line in the conversational prompt.
How to test against injection
In a test environment, include deliberately misleading documents alongside ordinary reference material. If a hidden instruction is added to a page being summarized, does the agent attempt an out-of-scope tool call? Does it cite the wrong source, request confidential data, or construct an unauthorized destination? Evaluate more than the visible answer: inspect actual tool calls, denied permissions, and server audit logs. Repeat tests across languages and multistep workflows. Do not copy production secrets or real user records into adversarial fixtures. The success criterion is not that the model never writes an odd sentence. It is that critical permissions remain enforced even if the model is persuaded to propose an unsafe action.
A defensible production workflow
Separate source retrieval, authentication, parsing, model interpretation, and publication into distinct application layers. Preserve the original URL and the exact cited passage; label findings as supported facts, interpretations, or unresolved uncertainty. Before any persistent write, apply schema checks, authorization, unique record identity, and revision comparisons. Keeping uncertain results in a review-only state need not block every unrelated item in the queue. Model capability can be measured by what it investigates; system safety is measured by the damage a wrong recommendation can cause. Treat OWASP guidance as a starting point for continuous boundary testing rather than a one-time prompt filter. Clear permissions, auditability, and reversible actions matter more than an impressive demonstration of autonomous behavior.
Comments
No comments yet. Be the first to contribute.
Reading “OWASP 2025: Prompt injection, RAG belgeleri ve ajan yetkisinin sınırları” raised this question for me: I would like clarity on user consent and data privacy.
Editor note: Product claims and measured outcomes should be assessed separately. A small pilot with defined metrics is a useful first step.
We should also discuss maintenance cost. An easy rollout does not guarantee trouble-free upgrades.
I would start with a pilot and measurable success criteria rather than assuming a launch means proven value.
✎ Write a comment