AI-generated image: Layered isometric plates in cyan, purple, and magenta gradient on dark background
Image generated by AI

System Prompt Leakage Got a New Name and a Bigger Scope

A model cannot keep a secret it can see.

That’s the whole argument behind a change OWASP made this year. In its 2026 Top 10 for LLM Applications, published early August 2026, OWASP renamed System Prompt Leakage to Hidden Context Exposure and widened what the category covers. Sensitive Information Disclosure held second place in the same list, and the list itself weighs 75% expert vote against 25% incident data drawn from 6,639 logged incidents. That’s not a committee guessing at what sounds scary, it’s counting what actually happened.

Why the old name was too narrow

System Prompt Leakage framed the problem as one string to protect: don’t let your system prompt leak. So people defended one string. They obfuscated the prompt, added a line telling the model not to reveal its instructions, maybe bolted on an output filter that greps for suspicious phrasing. Determined prompting got past most of that anyway, because the defence lived in the same channel as the attack, plain text arguing with plain text.

Hidden Context Exposure covers the actual shape of the problem: anything the model can see that the current user was never meant to see, regardless of which door it came in through. That’s also why the incident data matters here more than in most categories on the list. 6,639 logged incidents is a large enough sample that the pattern isn’t “occasionally a clever user extracts a prompt,” it’s “this happens across every surface a model reads from, routinely, whenever nobody drew a boundary around what goes into context.”

Four places a secret leaves the building

The system prompt. The original case, still real. An internal pricing rule or an unreleased feature name sitting in the instructions gets extracted by a user who asks the model to summarize everything above a certain point, or to translate its own instructions into another language.

Retrieved documents. A RAG index that includes an internal wiki page, pasted in for helpful context, with a partner’s contract terms still in it. It surfaces verbatim to the wrong user’s query because the index was never scoped to who’s allowed to see what.

Tool outputs. A tool call returns a full customer record because that was the fastest way to write the integration, and the model relays whatever’s in the response even though the interface only needed a name and an order status.

Memory. A model that carries context across turns, or across sessions, drags something over a boundary it shouldn’t. A detail from one customer’s support thread surfaces while answering a different customer, because the memory store was never scoped per tenant in the first place.

The four surfaces share a mechanism, not just a symptom. In every case, the model treated something as available context because a system upstream of it handed that content over without checking who was asking. The model isn’t the part that failed. The part that failed is whatever decided the content belonged in this particular conversation.

The defence that actually works

Not “tell the model to keep it secret.” A prompt instruction is advice, not a permission boundary, and nothing stops a user from asking the same question a hundred different ways until one of them works.

The actual fix sits below the model entirely: never let anything reach the model’s context that the current user isn’t already allowed to see. Enforce that at the data layer, the retrieval query that only returns documents scoped to this user, the tool response shaped to include only the fields the task needs, the memory store partitioned per tenant before a single token gets assembled into a prompt. Filter access first. Asking the model nicely afterward is not a control, it’s a suggestion the model happens to usually follow.

This is the same access-control discipline a web application already needs for its database queries and its API responses, just applied one layer further up, at whatever assembles the prompt. If your retrieval layer already scopes queries by user, and your tool responses already trim fields by permission, the model never sees the thing it could otherwise leak. The control has to exist before the prompt is built, not as an instruction inside it.

A static scanner catches the first of these four surfaces well. AIVory Guard can flag a secret pasted directly into a system prompt sitting in your repo, that’s a pattern in text it can see at scan time. It has no way to see a secret that arrives at runtime from a retrieval index you populated last month, or a tool response shaped wrong in production. That one needs a control at the data layer, not in the IDE.

I can list four leak surfaces because those are the ones counted in this year’s list. Do you actually know everything that reaches your model’s context right now, or are you assuming you do?