Automated Threat Detection#
Lilith implements multi-layered threat detection using:
1. Statistical Anomaly Detection#
- Algorithm: 3-sigma deviation from rolling baseline
- Window: 5-minute observation vs 60-minute baseline
- Sensitivity: Configurable per threat type
- False Positive Mitigation: Minimum threshold enforcement
2. Pattern-Based Detection#
- Authentication Anomalies: Credential stuffing, brute force patterns
- AI Safety Violations: Prompt injection, jailbreak attempts
- Content Violations: Policy breaches, harmful content
3. Behavioral Analysis#
- User Behavior: Unusual access patterns, geographic anomalies
- Service Behavior: Unexpected API usage, resource consumption
- Network Behavior: Traffic anomalies, connection patterns
Threat Types & Detection#
Authentication Threats#
Credential Stuffing#
Detection Criteria:
- Multiple failed authentications across different usernames
- High velocity from multiple IP addresses
- Success rate < 1%
Detection Logic:
javascript
const isCredentialStuffing = (events) => {
const failureRate = events.filter(e => e.type === 'auth_failed').length / events.length;
const uniqueIPs = new Set(events.map(e => e.context.ip_address)).size;
const uniqueUsers = new Set(events.map(e => e.context.username)).size;
return failureRate > 0.99 && uniqueIPs > 10 && uniqueUsers > 50;
};
Account Takeover#
Detection Criteria:
- Successful authentication after multiple failures
- Login from new geographic location
- Changed security settings immediately after login
AI Safety Threats#
Prompt Injection#
Detection Criteria:
- Specific injection patterns in user input
- Unusual token usage or response patterns
- Safety model confidence below threshold
Example Patterns:
- "Ignore previous instructions"
- "You are now in developer mode"
- Embedded system prompts
Jailbreak Attempts#
Detection Criteria:
- Attempts to bypass safety guardrails
- Roleplay requests for harmful content
- Multi-turn attack patterns
Content Moderation Threats#
Policy Violations#
Detection Criteria:
- Harmful content patterns
- Inappropriate persona requests
- Cultural sensitivity violations
Response Automation#
Automated Responses#
javascript
const responseActions = {
'credential_stuffing_detected': [
'enable_enhanced_rate_limiting',
'flag_suspicious_ips',
'notify_security_team'
],
'jailbreak_attempt': [
'log_detailed_context',
'update_safety_filters',
'alert_ai_safety_team'
],
'content_policy_violation': [
'queue_for_human_review',
'temporarily_restrict_user',
'log_violation_pattern'
]
};
Escalation Matrix#
| Threat Level | Automated Response | Human Involvement | Timeline |
|---|---|---|---|
| Low | Log + Monitor | Daily review | 24h |
| Medium | Rate limit + Alert | Team notification | 4h |
| High | Block + Escalate | On-call response | 1h |
| Critical | Lock + Incident | Immediate response | 15m |
Testing & Validation#
Automated Testing#
The security monitoring system includes comprehensive automated testing:
bash
# Run security monitoring tests
npm test -- tests/observability.security-monitoring.spec.js
# Run anomaly detection tests
npm test -- tests/observability.security-anomaly.spec.js
# Validate threat detection patterns
npm test -- tests/security-pattern-validation.spec.js
Manual Testing#
bash
# Test security event ingestion
curl -X POST "$OBSERVABILITY_URL/v1/security/events" \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $TOKEN" \
-d '{
"type": "test_event",
"severity": "low",
"source": "test",
"context": {"test": true}
}'
# Verify dashboard aggregation
curl "$OBSERVABILITY_URL/v1/security/dashboard"
Generated automatically by security-docs-generator.js Last updated: 2025-09-27T18:05:41.726Z