Skip to main content

What is a Signal?

A signal is the result of a compliance audit. It contains:
  • Risk Level: LOW, MEDIUM, HIGH, or CRITICAL
  • Violations: List of violation codes
  • Reasoning: Explanation of what was detected
  • Metadata: Provider, model, tokens, duration

Signal Structure

Risk Levels

LOW

Minor issues that don’t require immediate action:
  • Borderline content
  • Low-confidence detections
  • Informational warnings
Example: Ambiguous language that could be interpreted as biased

MEDIUM

Issues that should be reviewed:
  • Moderate policy violations
  • Potential PII leakage
  • Questionable content
Example: Email address in response without explicit user consent

HIGH

Serious violations requiring attention:
  • Clear PII leakage
  • Discriminatory content
  • Security vulnerabilities
  • Prompt injection attempts
Example: SSN or credit card number in response

CRITICAL

Severe violations requiring immediate action:
  • CSAM (zero tolerance)
  • Explicit dangerous instructions
  • Severe security breaches
  • Regulatory violations
Example: Any content involving minors in sexual context

Violation Codes

Signals include specific violation codes:

PII Violations

  • PII_LEAK: General PII detected
  • EMAIL: Email address
  • SSN: Social Security Number
  • CREDIT_CARD: Credit card number
  • PHONE: Phone number
  • PASSPORT: Passport number
  • HEALTH_DATA: Medical information
  • FINANCIAL_DATA: Financial information

Bias Violations

  • RACIAL_BIAS: Racial discrimination
  • GENDER_BIAS: Gender discrimination
  • AGE_BIAS: Age discrimination
  • DISABILITY_BIAS: Disability discrimination
  • RELIGIOUS_BIAS: Religious discrimination
  • INTERSECTIONAL_BIAS: Multiple bias types

Security Violations

  • CODE_INJECTION: SQL, command, or code injection
  • SECRET_LEAK: API keys, passwords, tokens
  • XSS: Cross-site scripting
  • SSRF: Server-side request forgery
  • DANGEROUS_INSTRUCTIONS: Harmful how-to content

Prompt Injection

  • PROMPT_INJECTION: Direct injection attack
  • JAILBREAK_ATTEMPT: DAN, STAN, or similar
  • SYSTEM_PROMPT_EXTRACTION: Attempting to reveal system prompt
  • GOAL_HIJACKING: Redirecting model behavior

Agent Safety

  • AGENT_LOOP: Infinite loop or recursion
  • IRREVERSIBLE_ACTION: Dangerous action without confirmation
  • SCOPE_CREEP: Acting outside intended scope
  • PRIVILEGE_ESCALATION: Attempting elevated permissions

Content Policy

  • VIOLENT_CONTENT: Graphic violence
  • SEXUAL_CONTENT: Sexual content
  • CSAM: Child sexual abuse material (CRITICAL)
  • HATE_SPEECH: Dehumanizing language
  • SELF_HARM_FACILITATION: Suicide or self-harm content

Viewing Signals

Dashboard

View signals in the dashboard:
  • Risk level breakdown
  • Recent signals
  • Filter by sandbox, provider, model
  • Date range filtering
  • Export for compliance reports

API

Query signals programmatically:
Response:

Filtering Signals

By Risk Level

By Sandbox

By Provider

By Date Range

Combined Filters

Signal Reasoning

Signals include redacted reasoning to protect PII:
This allows you to understand what was detected while protecting sensitive data.

Dashboard Statistics

Get aggregate statistics:
Response:

Exporting Signals

Export signals for compliance reports:
Formats: csv, json, xlsx

Webhooks (Coming Soon)

Receive real-time alerts for high-risk signals:

Best Practices

Monitoring

Set up regular monitoring:
  • Daily review of CRITICAL signals
  • Weekly review of HIGH signals
  • Monthly compliance reports

Alerting

Configure alerts for critical issues:
  • Email notifications for CRITICAL
  • Slack/Discord webhooks for HIGH
  • Dashboard monitoring for MEDIUM/LOW

Response Workflow

Establish a response workflow:
  1. CRITICAL: Immediate investigation and remediation
  2. HIGH: Review within 24 hours
  3. MEDIUM: Review within 1 week
  4. LOW: Monthly review

Compliance Reports

Generate regular compliance reports:
  • Monthly audit summaries
  • Violation trend analysis
  • Sandbox performance metrics
  • Regulatory compliance status

Next Steps

Dashboard

View signals in dashboard

API Reference

Query signals via API

Presets

Configure detection types with presets

Dashboard

View signals in the dashboard