AI Enable Kit
Developers
LLM08 · OWASP LLM Top 10 2026

Hidden Context Exposure

Strong by designRenamed and broadened from “System Prompt Leakage”

The invisible layer you place in front of the model — system instructions, tool definitions, retrieved documents, memory, configuration — leaks to users or attackers.

What it means

Hidden context is everything the application quietly puts in front of the model that the user never sees: the system instructions, the definitions of the tools it can call, documents pulled in behind the scenes, prior-turn memory, and configuration.

In 2026 OWASP renamed “System Prompt Leakage” to “Hidden Context Exposure” to cover this whole invisible layer, not just the opening instructions. If it leaks, it can reveal secrets, business logic, or a map of how to bypass your controls.

Why it matters

The rename widened the risk deliberately. Leaked context can expose embedded credentials, proprietary logic, or exactly the information an attacker needs to plan the next step.

How we approach it

Your system instructions and any custom skill definitions stay on the server and are never shipped to the browser — the client receives only what it needs to render the conversation.

Secrets are kept out of the prompt layer entirely: connection credentials are stored encrypted on the server, not placed into model context.

We minimize what enters context on each turn rather than accumulating it.

Honest caveat: like any LLM, a model can sometimes be coaxed into paraphrasing its own instructions. The durable defense — which we follow and recommend — is to never put a secret into an instruction in the first place.

Frequently asked

Can a user extract my system prompt?

The system prompt is never sent to the browser, which removes the easy path. No model is immune to being coaxed into paraphrasing its instructions, though — so the rule is to never place a secret in the prompt.

Where do custom skill instructions live?

On the server. Only the minimum needed to render the chat reaches the client; the instruction text itself stays server-side.

Should I put secrets or credentials in the prompt?

No. Keep them out of model context entirely. Credentials belong in encrypted server-side storage, not in instructions.

← LLM07MisinformationLLM09 →Vector and Embedding Weaknesses

All ten risks