Execution Stack

From input to execution to memory

Inspect how information becomes context, context becomes action, and action becomes durable state. Explore the attack surfaces, failure propagation, consequences, and countermeasures at every layer.

Keyboard: arrow keys move between layers. Motion is not required.

Memory is a feedback loop: durable state re-enters context on later runs.

Layer 01

Critical

Input

Accept user, document, web, email, API, and multimodal signals into the agent runtime before any model processing begins.

Viewing: Normal Operation

Purpose

Accept user, document, web, email, API, and multimodal signals into the agent runtime before any model processing begins.

Inputs

  • User messages and form fields
  • Documents, emails, tickets, and attachments
  • Web pages and third-party API payloads
  • Images, audio, and other multimodal content
  • Caller identity and channel metadata

Processing

  • Source identification and trust labeling
  • Content classification (instruction vs data)
  • Sanitization and isolation of untrusted payloads
  • Human confirmation gates for high-consequence requests

Outputs

  • Labeled input envelope
  • Provenance and channel attestation
  • Sanitized or quarantined content streams

Dependencies

  • Identity provider / channel auth
  • Content scanners and classifiers
  • Policy for trusted vs untrusted sources

Trust boundary

  • Untrusted external content enters a control plane that may treat text as instruction
  • Caller identity and source authenticity are asserted across systems
  • Multimodal payloads cross modality-specific parsers with uneven controls

Related chapters

Continue the investigation

Field manuals and related research

Use these chapters to deepen attack-path analysis, countermeasure design, and incident practice.

The IEG Field Manual is an AI security and countermeasures publication by BlockSiFr.