MANDATORY PREPARATION
Invoke /agent-workflow — it contains workflow principles, anti-patterns, and the Context Gathering Protocol. Follow the protocol before proceeding — if no workflow context exists yet, you MUST run /teach-maestro first.
Consult the guardrails-safety reference in the agent-workflow skill for the full defense-in-depth framework.
Add safety boundaries to a workflow. Guards protect against malicious inputs, unintended outputs, data leakage, cost explosion, and all the ways an autonomous system can go wrong in the real world.
Threat Assessment
Before adding guards, understand what you're protecting against:
| Threat | Risk Level | Guard Type |
|---|---|---|
| Prompt injection | High | Input sanitization, instruction hierarchy |
| PII leakage | High | Output filtering, data masking |
| Cost explosion | High | Token budgets, rate limits |
| Unauthorized actions | Medium | Permission scoping, confirmation gates |
| Hallucination | Medium | Source attribution, fact checking |
| Service abuse | Medium | Rate limiting, authentication |
Guard Implementation
Input Guards
textBefore processing any input: 1. Validate against schema (reject malformed) 2. Check size limits (reject oversized) 3. Sanitize for injection patterns 4. Rate limit check (reject if exceeded) 5. Authentication/authorization check
Output Guards
textBefore returning any output: 1. Schema validation (format correct?) 2. PII scan (names, emails, SSNs, etc.) 3. Content policy check 4. Confidence threshold check 5. Source attribution present?
Cost Guards
textBefore every model/API call: 1. Check remaining budget 2. Estimate request cost 3. If estimate > remaining budget → reject or use cheaper alternative 4. After call → update spent amount 5. Circuit breaker check (too many failures?)
Permission Guards
textFor every tool call: 1. Is this tool allowed for this user/context? 2. Is this a destructive operation? → require confirmation 3. Is this accessing data the user is authorized for? 4. Log the access for audit trail
Guard Checklist
- All inputs validated before processing
- PII detection on all outputs
- Cost ceiling set with enforcement
- Prompt injection defenses active
- Destructive operations require confirmation
- All access logged for audit
- Circuit breakers on external services
- Rate limits on all endpoints
Recommended Next Step
After adding guards, run /evaluate with adversarial test scenarios to verify guards hold under attack.
NEVER:
- Deploy without input validation
- Trust model output for high-stakes decisions without verification
- Run without cost controls
- Skip logging (you need the audit trail)
- Assume the model will follow safety instructions 100% of the time
