Skip to main content

Structural & Meta-Level Tactics

Techniques that exploit how the model processes context, instructions, or its own rules.

These go deeper than prompt-level tactics. Instead of manipulating what you ask, they manipulate how the model thinks about what you ask—targeting the reasoning process, the instruction hierarchy, or the safety mechanisms themselves.


Techniques

TechniqueWhat it does
ICL ExploitationManipulate in-context learning with crafted examples
Control Plane ConfusionBlur the line between system instructions and user input
Meta-Rule ManipulationTarget the model's understanding of its own constraints
Capability InversionTurn helpful capabilities against intended use
Cognitive LoadOverwhelm the model's attention or reasoning
Defense EvasionBypass safety classifiers and filters

When to use structural tactics

Use these when:

  • Prompt-level tactics are hitting consistent, well-trained refusals
  • The target has a system prompt you can probe or inject into
  • You want to attack the safety mechanism itself, not just evade it
  • The model uses few-shot examples or retrieval-augmented generation

Move to infrastructure tactics when:

  • The target is an agent with tool access
  • You can poison data sources the model consumes
  • The attack surface extends beyond the chat interface

Decision framework

Structural tactics require understanding the target's architecture. Before using these, complete target analysis to understand what context the model has access to.

For combining structural with prompt-level tactics, see Composition.