Execution stack

From input to memory

Select a layer to inspect purpose, trust boundaries, failure modes, consequences, and control opportunities. Keyboard: arrow keys move between layers. Motion is not required.

Layer 01

Input

Accept user, document, web, email, and API inputs into the agent runtime.

Trust boundary
Untrusted external content enters a control plane that may treat text as instruction.
Failure modes
  • Direct prompt injection
  • Channel spoofing
  • Attachment-based injection
Representative vulnerabilities
  • V001 Direct prompt injection
  • V002 Indirect prompt injection
  • V014 Attachment-based injection
Consequences
  • Instruction override
  • Policy bypass
  • Unsafe tool priming
Control opportunities
  • Input linting
  • Channel authentication
  • Content sanitization

Compromise at this layer can propagate forward into later layers—especially execution and memory—unless gated.