Runtime Forensics

Runtime forensics is the AISecOps v0.5 replay foundation for reconstructing how an agentic AI system reached a runtime decision. It connects prompts, skills, retrieved context, memory, tool results, policy checks, approvals, and execution outcomes into replayable evidence.

What Runtime Forensics Means

AISecOps reconstructs agent execution history so security and platform teams can understand what happened, why it happened, and which instruction source influenced the decision. The goal is not only observability; it is forensic reconstruction of autonomous runtime behavior.

Why Logs Are Not Enough

Traditional logs are event records. They can show that a tool was called, a request failed, or a policy returned a result. AISecOps replay logs are decision-chain evidence: they preserve the execution plan, provenance inputs, capability result, policy result, and final runtime decision in a structured JSONL format.

Replayable Audit Model

Prompt / Skill / Memory
        ↓
Execution Plan
        ↓
Capability Check
        ↓
Policy Evaluation
        ↓
Approval / Block / Execute
        ↓
Structured JSONL Audit
        ↓
Replay Engine

Each stage emits structured evidence that can be replayed without trusting the model to explain itself after the fact.

Instruction Provenance

AISecOps tracks instruction and context sources as first-class provenance fields. A replay trace can identify whether a decision was influenced by:

This distinction matters during incident response because policy violations often originate from indirect instructions rather than the original user request.

Replay CLI

The replay workflow starts from a trace identifier and the JSONL audit file produced by the runtime control plane. 

aisecops-replay --trace-id run-123 --audit-file audit/events.jsonl

Example Replay Output

trace_id: run-123
provenance:
  user_prompt: "summarize the vendor renewal"
  retrieval_chunk: contract-renewal-2026.md#section-4
  skill: vendor_review
  tool_result: crm.lookup_vendor
execution_plan_id: plan-8f42
capability_result:
  capability: cap_contract_review
  status: allowed
policy_result:
  policy: external_email_recipients
  status: approval_required
  reason: recipient_not_allowlisted
final_decision: block_pending_approval

The replay output is designed for investigation workflows, not model-generated justification. It gives responders the runtime facts needed to reconstruct the decision boundary.

Enterprise Use Cases

Current Limits

Roadmap