The rise of autonomous AI agents has created a blind spot where legitimate credentials are used for unintended actions, evading traditional detection.
AI IR Overlay provides a structured framework to address this gap by focusing on agent capability mapping and intent analysis rather than credential hygiene.
Central to the framework is the AI Bill of Materials (AI-BOM), a YAML artifact that records the deployed identity, granted scopes, tools, write-enabled targets, memory, and data sources of each production agent.
Unlike design-time diagrams, the AI-BOM reflects the actual runtime capabilities, enabling teams to quantify blast radii before any model inspection.
The framework introduces a graduated containment ladder from M0 to M5, with M1 stripping write-capable tools while preserving read-only functions for forensic preservation.
This containment step maintains the agent’s internal state, allowing responders to collect prompts, tool calls, and memory snapshots without destroying evidence.
Complementing the ladder is a minimum evidence set comprising six classes: prompts/responses, tool-call logs, retrieval traces, internal memory, configuration snapshots, and relevant identity/audit logs.
Capturing these six evidence types ensures that even stealthy, policy-compliant misuse leaves a traceable footprint for post-mortem reconstruction.
A validation helper (scripts/verify.py) checks JSON-schema conformity and data freshness, operating safely in CI pipelines without invoking the agent itself.
The actuation layer defines a kill-switch API contract and reference implementations that demonstrate moving an agent through containment states M0-M4.
Notably, the specification omits autonomous trigger logic, assuming a staffed SOC can issue containment orders within ten minutes for Tier-1 incidents.
Organizations must therefore decide how to bridge the gap when human judgment is unavailable, either by strengthening SOC coverage or adding automated escalation policies.