| name | Remediation Playbook |
| description | Actionable fix patterns for LLM vulnerabilities — system prompt hardening, input validation, output filtering, and tool sandboxing |
| version | 1.0.0 |
Remediation Playbook
Purpose
Provides the remediation-agent with concrete fix patterns for each vulnerability class. Every fix should be specific, implementable, and verifiable.
When to Use
- After scan results show vulnerabilities
- When the user asks "how do I fix this?"
- When generating remediation reports
- During security reviews
Fix Categories
| Category | Fixes | When to Apply |
|---|
| System Prompt Hardening | Instruction boundaries, explicit denials, XML-tagged input isolation | Jailbreak, injection findings |
| Input Validation | Length limits, character filtering, injection detection, sanitization | Injection, tool abuse findings |
| Output Filtering | PII redaction, content moderation, response validation | PII, harmful content findings |
| Tool Sandboxing | Parameterized queries, permission scoping, allowlists, capability limits | Tool abuse findings |
Priority Framework
- Critical — Fix immediately: prompt injection with data access, PII exposure, tool abuse with DB/file access
- High — Fix this sprint: jailbreaks bypassing core restrictions, bias in customer-facing apps, compliance violations
- Medium — Fix next sprint: partial prompt leaks, edge case harmful content, low-confidence findings
- Low — Track and monitor: cosmetic issues, theoretical attacks, defense-in-depth improvements
Reference Files
