ContentsThe library

What a Model Refuses, and Why

The Attacker Does Not Have to Be the One Typing

Last timeChecking What Goes Out

If your system reads documents, pages or tool output, somebody who can write into any of those is writing into your prompt. Here is the path, step by step.

Every door into the prompt

The previous lessons assumed the attacker is the person typing. Usually they

are not, and the reason is that very little of a modern prompt is typed by the

user.

Count the sources. The system prompt, which you wrote. Retrieved documents from

your own index. Web pages fetched by a tool. The result of an API call. A file

the user uploaded. Earlier turns of this conversation. Possibly a summary of

earlier turns, written by the model.

Every one of those except the first is text somebody else produced, and each of

them is a surface somebody can write to.

FIG 1Who can write into each source
writeable by someone outreviewed before it is intreated as untrusted by share of prompt tokens,
system prompt0110
internal knowledge base10012
customer support tickets10040
web pages fetched by a t10095
uploaded files1108
results from a partner A10030
The marked row is the worst case: a web page is writeable by literally anyone, is not reviewed, is not treated as untrusted, and can fill most of the context. The second column is almost entirely zeros, which is the finding. Very little of what reaches the model has been looked at by anybody.

The knowledge base deserves a word. It feels internal and safe. It usually

contains content written by support staff, pasted from customer emails, synced

from a wiki anybody in the company can edit, and sometimes imported from a

public source years ago. Internal means inside the company, not written by

somebody you have vetted for this purpose.

Tracing one through

The exercise worth doing is following a single attack from where it was written

to where it had an effect. Nobody believes the attack until they write out the

path for their own system.

The lesson stops here

3 more paragraphs to go

You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.

The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.

See the planThe contents

This is the reading half

Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.

The contents

The rest of this course

  1. 01A Habit, a Guard, and a Note on the Door
  2. 02Your Rules and Their Attack Arrive in the Same Envelopeopening only
  3. 03Two Mistakes, One Dial, and You Have to Chooseopening only
  4. 04The Only Layer That Sees What You Are About to Sayopening only
  5. 05The Attacker Does Not Have to Be the One Typingyou are here
  6. 06Build It So the Worst Case Is Boringopening only
  7. 07Nobody Writes In to Report a Refusalopening only
  8. 08If Two People Cannot Apply It, It Is Not a Ruleopening only

Read alongside