Guides & analysis

When your AI reads a webpage, who gets to give it orders?

A document can contain instructions the user never authorized. OWASP's guidance explains why keeping external content separate, limiting tool powers and checking consequential actions are different protections. This guide offers reader questions, not a product security certification.

By SpicyRanksPublished 2 linked sources ↓

A source can try to become the boss

OWASP describes indirect prompt injection as external material, such as a website or file, influencing a model's behavior in unintended ways. It identifies possible consequences including manipulated output, disclosure and unauthorized use of connected functions. These are general risks, not allegations about a particular service. [1]

Spicy's takeAnalysis: imagine asking an assistant to summarize a supplier's brochure. If the brochure also tells it to send an internal file somewhere else, that second request did not come from you. The important question is whose authority the proposed action rests on, even if the instruction arrives inside relevant reading material. [1]

Separate advice from the power to act

OWASP's prevention guidance recommends restricting privileges, identifying external content and using human approval for high-risk operations. Its cheat sheet adds that labels and prompt templates do not themselves enforce a boundary; proposed tool arguments and permissions need checks outside the model. [1][2]

Spicy's takeAnalysis: before connecting a service, ask whether the assistant needs read access, write access or both. A reading task is a good reason to question an unrelated sending permission. Where an action has consequences, a useful preview should show the destination and intended change clearly enough for you to evaluate it. [2]

A reassuring sentence is not a test result

The OWASP cheat sheet says keyword screening can miss rephrased instructions and can block innocent requests. It treats guardrail models as additional layers, alongside deterministic controls. Its sample smoke tests are illustrative rather than a representative security benchmark. [2]

Spicy's takeAnalysis: ask a vendor what was tested, which actions were protected and what happens when a check fails. “We block malicious prompts” leaves those questions open. Prefer a concrete explanation of enforced permissions and failure behavior to a demonstration that one suspicious sentence was refused. [2]

Spicy's takeAnalysis: the goal is proportionate authority, not constant interruption. Give an assistant the scope needed for the task, understand what it can change and inspect meaningful pending actions when review is available. This guide does not test any app or certify its defenses; it explains what to ask before trusting a connection. [1][2]

Sources & further reading

Source material checked Nov 20, 2026. Linked material may change after publication.

  1. LLM01:2025 Prompt Injection ↗

    OWASP GenAI Security Project · Primary source

  2. LLM Prompt Injection Prevention Cheat Sheet ↗

    OWASP Cheat Sheet Series · Primary source