AI Pilot — Advanced AI Audit

Audience: Compliance, security architects, audit owners Time: ~8 min read

AI Pilot ships the deepest audit trail in the industry for AI traffic: every request carries decision lineage, prompt and response hashes, redaction history, and tool-call chains. Events flow through Control Core's existing SIEM outbox to Splunk, Elastic, QRadar, Sentinel, or Grafana Loki.

What is captured

Event	When	Key fields
`AI_TRAFFIC_LOG`	Every AI request	route, provider, model, decision, status, tokens_in, tokens_out, cost_usd
`AI_DECISION_LINEAGE`	Every decision	rule chain, OPA policy IDs, principal, tenant, environment
`AI_PII_REDACTION`	Redaction event	detector, count, hashed location, action
`AI_POLICY_VIOLATION`	Deny	rule that denied, principal, requested target
`AI_MCP_TOOL_CALL`	MCP tool invocation	tool, args_hash, server, decision, cache_hit
`AI_PROVIDER_TOKEN_ROTATED`	Upstream auth rotation	provider, old/new token id, reason
`AI_BREAKER_TRIPPED`	Circuit breaker	signal, value, threshold, fallback used
`AI_FALLBACK_USED`	Fallback fired	from, to, reason
`AI_CACHE_HIT` / `AI_CACHE_MISS`	Cache lookup	key fingerprint, semantic match, ttl

Every event includes a stable correlation_id so you can stitch the chain end-to-end in your SIEM.

Configure

Open Settings -> AI Pilot -> Advanced Audit:

Field	Description
Enabled	Master toggle
Sampling	`100% / 50% / 10% / 1%` for high-volume environments
Capture prompt hash	SHA-256 of the prompt body — never the raw text
Capture response hash	SHA-256 of the response body
Capture tool-call chain	Full JSON-RPC tool chain for MCP traffic
Redaction log	Per-detector counts and locations
SIEM target	Reuses the platform-wide SIEM target from `Settings -> Audit logs -> SIEM`
Retention (days)	How long to keep events in the AuditLog table before TTL

Sampling cuts log volume; lineage and policy events always sample at 100% regardless of the body sampling rate.

Decision lineage

Every AI decision records the rule chain that produced it:

{
  "event": "AI_DECISION_LINEAGE",
  "correlation_id": "8c7e...",
  "principal": {"sub": "u_123", "role": "finance_manager", "mfa": true},
  "target": {"type": "llm", "provider": "openai", "model": "gpt-4o"},
  "decision": "allow",
  "rule_chain": [
    {"package": "controlcore.ai.access", "rule": "allow_finance_managers", "matched": true},
    {"package": "controlcore.ai.cost",   "rule": "under_daily_cap",          "matched": true},
    {"package": "controlcore.ai.safety", "rule": "no_pii_in_prompt",         "matched": true}
  ],
  "redaction_count": 2,
  "tokens_in": 412,
  "tokens_out": 893,
  "cost_usd": 0.024
}

This lets compliance reviewers reconstruct why a request was allowed/denied, not just what.

Prompt and response hashing

The bouncer stores SHA-256 hashes of the request and response bodies (and optionally a redacted preview, configurable). Hashes let you:

Verify whether two requests had identical content without storing the content
Cross-reference cache hits with audit rows
Build dedupe metrics for prompt engineering work

Raw prompt/response bodies are never captured by default. Enable previews only in Sandbox or with explicit DPO sign-off.

SSE live audit stream

The /pilot -> Audit tab opens an SSE connection to GET /pep-config/pilot/bouncer/{id}/audit-stream and renders a live trace. You can:

Pause / resume the stream
Drill into any row to see the full lineage
Deep-link any event to /audit?correlation_id=... for SIEM-equivalent forensics

SIEM wiring

Audit events flow through the existing SIEMOutboxEntry durable outbox. Configure the destination once at Settings -> Audit logs -> SIEM; every AI event gets the same delivery guarantee as the rest of Control Core's audit traffic.

The Pilot dashboard's "SIEM outbox (AI)" card shows pending / delivered / failed counts and the last delivery timestamp.

Compliance use cases

Standard	What this enables
GDPR Article 22	Show the rule chain that produced each automated decision
HIPAA	Prove PHI redaction happened before any provider call
SOC 2 / ISO 27001	Continuous evidence of policy enforcement on AI traffic
FINTRAC / FinCEN	Tie AI-assisted decisions to a deterministic, audited rule chain
EU AI Act	Records of automated decisions with traceability