Domain 6 of 6

Running It in Production

Everything that starts once real users arrive. Instructions that arrive inside the data, the guards around an untrusted output, shipping a prompt change safely, watching quality rather than only errors, what happens when the provider changes the model underneath, and the practice that keeps all of it from decaying.

Concepts

~15%

Of the exam

Practice questions

Concepts in this domain

01Prompt injectionText that reaches a model as data and gets acted on as instruction. The direct and indirect forms and why they are different problems, the request as the model actually receives it, the three conditions that turn an injection into a breach, and why the defence that works assumes the attack succeeds.02Defending a language model featureContainment for an attack that cannot be prevented. Treating model output as untrusted input, writing tool scopes down as a policy file, approval gates on consequential actions, and the six published patterns that constrain an agent once it has read something a stranger wrote.03Versioning and deploying promptsA prompt as a deployable artefact with a version, an owner and a changelog. The registry entry a request resolves against, sticky assignment to a share of traffic, the model version held constant while the prompt moves, and the conditions that stop a rollout without anybody being paged.04Monitoring a language model featureWatching quality rather than only errors, because this feature fails by returning something plausible. A stratified sample of live traffic scored by the validated judge, the signals that need no judge at all, and a burn rate alert adapted for a measurement that arrives a few hundred samples an hour.05Model migrationWhat happens when a provider retires a model, updates one behind an alias or prices one differently. The published notice periods, the difference between a dated version and an alias, why a prompt tuned for one model does not transfer, and the comparison that decides whether a candidate is ready.06Building a prompt and context practiceWhat a team needs to do this repeatedly. A prompt registry, an eval set that grows from production, a weekly habit of reading traces and a written record of what was tried, with the rhythm that keeps all four alive and the one idea every technique in the course comes back to.

Try a question from this domain

A facilities management company summarises inbound tickets for a weekly report. The request its application sends is below, and the assistant replied with the single word PINEAPPLE. The proposed fix is a firmer system prompt telling the model to ignore instructions inside tickets. What does a stronger assembly give that the firmer prompt does not?

AIt moves the team instruction after the ticket, so that the most recent instruction in the sequence is always the one the team wrote.
BIt enforces the boundary at the level of the interface, in the way a parameterised query separates a command from the values it carries.
CIt removes the attack before the model reads it, since a labelled block is passed through a filter that a plain message is not.
DIt gives the untrusted material a boundary somebody can point at, states the rule about that boundary before the material arrives, fixes the output shape and defines what the feature does when an attack is present.

13 questions on this domain.

One per page, with a worked explanation.

Start the set