Skip to the content.

Containment is not one switch. Six modes with TTA targets.

This file is one self-contained piece of the AI IR Overlay™ framework. Cross-references to other pieces point to other packages in the same set, which you can obtain at jacobideji.com.


Kill-Switch Modes: The Containment Ladder

The “kill switch” is not one switch. It is six modes that escalate from passive observation to full disable, and back.

Binary on/off is rarely appropriate in production. The Overlay defines six modes so containment can be calibrated to confidence, impact, and business need.


The Modes at a Glance

Mode Name Use when TTA target Approver
M0 Observe Normal operations n/a Owner
M1 Read-Only Suspicious behavior; low/moderate impact ≤ 10 min Tier-1 SOC
M2 Approvals Required Agent must keep operating; actions need two-person rule ≤ 10 min Tier-1 SOC
M3 Tool Tiering Targeted containment. Disable high-risk tools only ≤ 10 min Tier-1 SOC
M4 Full Disable Active harm, confirmed misuse, or evidence of compromise ≤ 10 min Tier-1 SOC
M5 Controlled Re-Enable Containment validated; staged recovery n/a CISO / IC

TTA = Time-To-Activate. Measured from incident-commander order to mode in effect.


Mode 0: Observe (Baseline)

Purpose: Normal operations with logging.

Requirements:

Exit criteria: Incident declared. Step up to M1 (or further) based on confidence.


Mode 1: Read-Only (Preferred First Containment)

Purpose: Stop writes without stopping the business.

What changes: All write tools are stripped from the agent’s tool set. Read and query tools remain.

Use when:

Operational checks:

Exit criteria: Investigation confirms benign (return to M0), confirms harm (step up to M3/M4), or needs continued operation with control (step to M2).


Mode 2: Approvals Required (Two-Person Rule)

Purpose: Continue operation, but no action without a human approver.

What changes: Every tool call is queued for human approval before execution.

Use when:

Operational checks:

Exit criteria: Risk reduced (return to M1 or M0), or containment must escalate (M3/M4).


Mode 3: Tool Tiering

Purpose: Contain selectively. Disable high-risk tools, keep low-risk.

What changes: Specific tools (external email send, code deploy, financial actions) are disabled. Lower-risk tools (internal search, status lookup) remain.

Use when:

Operational checks:

Exit criteria: Vector contained (step down to M1/M2), or scope expands (M4).

RAG-specific containment: when the suspected attack path is the retrieval layer (RAG, knowledge base, vector index), use the M3-RAG variant. Cut retrieval to the suspect corpus and leave the other corpora alone. The agent keeps working with reduced knowledge but no exposure to poisoned content. See Playbook 03: RAG / Knowledge-Base Forensics for the freeze-the-world sequence and the seven-component pipeline forensics.


Mode 4: Full Disable

Purpose: Hard stop.

What changes: Agent is taken offline. Active sessions are terminated. Tokens are scoped for revocation (not yet rotated; see Evidence Plan).

Use when:

Operational checks (critical sequence):

Rotating tokens before capturing scopes is the single most common evidence-destruction failure in AI IR.

Exit criteria: Eradication complete. Move to M5 controlled re-enable.


Mode 5: Controlled Re-Enable (Recovery)

Purpose: Restore operation in stages, with validation at each step.

Sequence:

  1. Re-enable in Read-Only (M1). Confirm the agent functions and logs flow.
  2. Validate retrieval and tool policies. Corpora versions confirmed clean.
  3. Replay the incident scenario in a safe harness. Confirm fix holds.
  4. Re-enable tools incrementally. Start with low-risk, monitor for drift.
  5. Return to M0 Observe. Only after all of the above.

Approver: CISO or designated Incident Commander. Never the original agent owner alone.

Operational checks:


Pre-Production Requirements

Before ANY agent reaches production:

If any of the above is “not yet,” the agent is not production-ready.



Source: AI IR Overlay newsletter and framework synthesis, by Jacob Ideji. https://www.linkedin.com/in/jacobideji/