Security Governance

Definition

Security governance is the preflight layer that detects sensitive input, masks PII, blocks prompt injection, and writes audit records without storing raw sensitive text.

Why It Exists In Aurelia Ledger

Financial enterprise AI must handle confidential data and adversarial prompts before retrieval or tool use. Unsafe requests should not reach downstream agents.

Implementation Links

Area	File	Lines	Why It Matters
Detection patterns	security_service.py	L10-L27	Defines PII and secret-like patterns
Security check entrypoint	security_service.py	L28-L71	Runs mask/block logic and audit logging
Block response and hashing	security_service.py	L72-L86	Builds governance response and hashes messages
PII and injection detection	security_service.py	L87-L127	Masks PII and detects prompt injection
Action and risk mapping	security_service.py	L128-L147	Maps findings to allow, mask, or block
Audit table	models.py	L50-L62	Stores message hash and governance metadata

Core Workflow

flowchart TD
    Message[User Message] --> Detect[Detect PII / Injection]
    Detect --> Action[Decide Action]
    Action -->|allow| Route[Continue To Orchestrator]
    Action -->|mask| Mask[Use Masked Message]
    Mask --> Route
    Action -->|block| Block[Governance Response]
    Detect --> Audit[(Security Audit)]

Technical Deep Dive

The security layer is deterministic. It does not call an LLM moderation API. This keeps latency low and behavior predictable.

The system masks medium-risk sensitive data and blocks high-risk prompt-injection patterns. Audit records store message hashes, not raw sensitive messages.

Formula / Scoring Model

Risk-action mapping:

if prompt_injection_detected:
    action = block
elif pii_detected:
    action = mask
else:
    action = allow

Message hash:

message_hash = SHA256(raw_message)

Audit record:

audit = { message_hash, risk_level, action, finding_count, agent, created_at }

Example Walkthrough

Input:

Contact analyst at test@example.com about Apple risk

Expected behavior:

1. Email pattern is detected.
2. Risk level becomes medium.
3. Action becomes mask.
4. Message becomes Contact analyst at [EMAIL] about Apple risk.
5. Masked message continues to routing.

Design Tradeoffs

• Rule-based detection is transparent and fast.
• It does not replace enterprise DLP.
• False positives and false negatives are possible.

Failure Modes

• Sophisticated prompt injection may bypass patterns.
• Legitimate text may be blocked if it resembles a secret.
• Role-based permissions are not fully implemented yet.

Exercises

1. Checkpoint: Explain why security preflight should run before retrieval.

2. Hands-on: Inspect security_service.py L128-L147 and describe the action decision logic.

3. Interview Drill: Explain why audit logs store message hashes instead of raw messages.

Interview Explanation

Security is part of the workflow, not a separate afterthought. The platform checks input before any downstream agent or retrieval tool can act on it.