Skip to content
DarkPrint
Risk marker

Prompt injection

autogen/prompt-injection

The node accepts text from outside that may carry instructions aimed at the model reading it.

Where it sits

1 level deep
Broader chain
  1. isolation-breach
  2. autogen/prompt-injection

Each term is a kind of the one before it. A rule written about any term in this chain also applies to autogen/prompt-injection.

Narrower terms

No term is narrower than autogen/prompt-injection.

What it costs

declared on the card
41.50points, once per blueprint

A blueprint starts at a clean 4, loses the weight of every marker present, and the result is clamped into 1–4. autogen/prompt-injection is charged once for the whole blueprint no matter how many nodes carry it. The explanation still lists every node that established it.

Nothing in the topology can establish autogen/prompt-injection on its own, it is a fact about what the node does that only its author can state. The analyzer takes the card at its word and names the node in the explanation.

The engine's configuration prices the curated markers. It is silent about this one. The number is the namespaced term's own declared weight, read from the vocabulary the bundle ships. Nothing records which weight an evaluation was made under, so move it and two evaluations stop being comparable.

Who uses it

2 cards · 1 blueprint