Written for Security, technology, risk and operations leaders who have to put a large-language-model system into production, and the sponsors who have to fund one.
What's in this paper
This 21-page paper is a design framework for putting a large-language-model system into production. The library has sometimes described it as a third-party “Top 10,” a red-team methodology, and an attack taxonomy. It is not those things. It will not paste a third-party list as if it were ours, name other firms, or provide an exploit, a payload, or a step-by-step attack procedure.
What it is, instead, is twelve tests to apply before a first case is funded, and to the systems that are already running; eight threat classes, described qualitatively, so that T1 can be answered honestly; and a way to treat instruction override, data leaving, insecure output, over-reach, poisoned material, unsanctioned use, vendors, and access as architecture. Passing the tests makes a failure visible, traceable and stoppable. It does not make it impossible. This paper will not certify that a system is secure.
It is written for security, technology, risk and operations leaders, and for the sponsors who have to fund the work. The illustrative community bank in the paper is the same hypothetical firm as in the 2026 diagnostic. It is labeled illustrative, and is not a client case study. A companion workbook holds the tests and a threat-class sheet. A completed workbook is not a finding that a system is secure.
How the tests are meant to be used
A failed T1, T2, T3, T4, T5 or T9 is usually a reason to pause. Threats are named for this system, not borrowed from a ranking. What would leave — prompts, completions, logs, retrieval, evaluation material, fine-tune material, vendor paths — is mapped. Instruction override is treated as a design: defenses are architecture, not a prompt. A draft is not a trusted action. Tools and write-backs are gated; the honest first case is often read-only, with no write-back.
Corpus and evaluation material are named, approved, and not quietly edited. Access is named and no broader than the work; there is no shared “the model” account, and permissions travel with a retrieved passage. Vendors and the model path are in view — the vendor scorecard in this library is the method, and a hosted path described as “private” is still a path. A named operator who is not only the builder, a switch-off, and a trail that can reconstruct a day, are in the first design.
Unsanctioned use is on the inventory, or is refused; an honest approved path is part of the design. Training and reuse are written down. Security is re-measured after a change: a new model, tool, corpus or vendor is a new measure, and a passing score is not a certification. The RAG guide, the governance framework, the vendor scorecard and the offline guide in this library are companions.
Twelve design tests (excerpt)
The same standard for a drafting assistant and for a system that can call a tool. The answers will differ. The standard does not.
T1
The threat classes that apply are named. Which of C1 through C8 apply, written down, for this system. Not a third-party ranking.
T2
What would leave is mapped: prompts, completions, logs, retrieval, evaluation, fine-tune material, vendor paths.
T3
Instruction override is treated as a design. Direct, and through material the system is asked to read. Defenses are architecture, not a prompt.
T4
A draft is not a trusted action. A completion is not a message, a record, or a command until a person or a deterministic check has stood behind it.
T5
Tools and write-backs are gated. Least privilege. The honest first case is often read-only, with no write-back.
T6
Corpus and evaluation material are approved: named, approved, and not quietly edited.
T9
An operator, a switch-off, and an audit trail. A named operator who is not only the builder, and a trail that can reconstruct a day.
T12
Security is re-measured after a change. A new model, tool, corpus or vendor is a new measure. A passing score is not a certification.
The remaining tests cover named access, vendors and the model path, unsanctioned use on the inventory, and training and reuse written down. The complete twelve are in the PDF and the workbook.
Eight threat classes, qualitative
T1 asks which of these classes apply to this system. This paper describes the class. It will not show you how to do it, and it will not give you a string to try.
C1
Instruction override — a prompt, or material the system is asked to read, that tries to override the system’s instructions. Direct, and indirect, through content the system is given.
C2
Data leaving — prompts, completions, logs, retrieved passages, evaluation sets, fine-tune material, and anything a vendor would see. If it is not mapped, T2 has not been finished.
C3
Poisoned material — a corpus, an evaluation set, or fine-tune material that was not approved, or that was quietly changed. The class is integrity of what the system is taught and measured against.
C4
Over-reach — a system that can call a tool, write back, or act, is steered toward something it should not do. The agentic-workflows paper is the companion for the gate.
C5
Insecure output — a draft treated as a trusted action: forwarded into a message, a record, a tool, or another system, without a person or a deterministic check.
C6
Unsanctioned use — tools and assistants people actually use that are not on the inventory. Shadow use is a governance failure as well as a security one.
C7
Vendor and model-path exposure — a hosted path, a subprocessor, or a “private” endpoint that is still a path. The vendor scorecard is the method.
C8
Access that is broader than the work — a shared account, a retrieved passage without the permission that should travel with it, or a trail that the wrong people can read.
Naming the classes is T1. Closing any one of them in isolation is not a close of the design.
Get the guide and workbook
Enter your name, work email and company. The files appear on this page immediately. A backup download link is also emailed when email delivery is available.