safety-guardrails

name	safety-guardrails
description	LLM safety guardrails and content moderation
sasmp_version	1.3.0
bonded_agent	08-prompt-security-agent
bond_type	PRIMARY_BOND

Safety Guardrails Skill

Bonded to: prompt-security-agent

---

Quick Start

Skill("custom-plugin-prompt-engineering:safety-guardrails")

---

Parameter Schema

parameters:
  safety_level:
    type: enum
    values: [permissive, standard, strict, maximum]
    default: standard

  content_filters:
    type: array
    values: [harmful, hate, violence, adult, pii]
    default: [harmful, hate, violence]

  output_validation:
    type: boolean
    default: true

---

Guardrail Types

Guardrail	Purpose	Implementation
Input filtering	Block harmful requests	Pattern matching
Output filtering	Prevent harmful outputs	Content analysis
Topic boundaries	Stay on-topic	Scope enforcement
Format validation	Ensure safe formats	Schema checking

---

Content Filtering

Categories

content_categories:
  harmful:
    - dangerous_activities
    - illegal_actions
    - self_harm

  hate_speech:
    - discrimination
    - slurs
    - targeted_harassment

  violence:
    - graphic_violence
    - threats
    - weapons_instructions

  pii:
    - personal_data
    - financial_info
    - credentials

Filter Implementation

## Content Guidelines

NEVER generate content that:
1. Provides instructions for harmful activities
2. Contains hate speech or discrimination
3. Describes graphic violence
4. Exposes personal information
5. Bypasses safety measures

If a request violates these guidelines:
1. Decline politely
2. Explain which guideline applies
3. Offer a safe alternative if possible

---

Output Validation

validation_rules:
  format_check:
    - valid_json_if_requested
    - no_executable_code_in_text
    - no_embedded_commands

  content_check:
    - no_pii_exposure
    - no_harmful_instructions
    - appropriate_for_audience

  consistency_check:
    - matches_role_constraints
    - within_topic_boundaries

---

Safe Response Patterns

Declining Harmful Requests

I can't help with that request because [reason].

Here's what I can help with instead:
- [Alternative 1]
- [Alternative 2]

Would any of these work for you?

Handling Edge Cases

I notice this request is [description of concern].

To ensure I'm being helpful in the right way:
1. Could you clarify [specific aspect]?
2. Here's a safe approach to [related task]:
   [Safe alternative]

---

Troubleshooting

Issue	Cause	Solution
Over-blocking	Too strict	Tune sensitivity
Under-blocking	Too permissive	Add patterns
False positives	Ambiguous content	Context-aware rules
Inconsistent	Rule conflicts	Prioritize rules

---

References

See: Anthropic Constitutional AI, OpenAI Moderation API