Custom Policies

Create custom policy files for GPT OSS Safeguard.

Policy File Format

Location: src/guardrails/policies/

Format: Plain text with {input} placeholder

Example Policy

src/guardrails/policies/my_policy.txt:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

You are a content safety validator for a family-friendly app.

Evaluate this user input against our policies:
1. No violence or graphic content
2. No profanity or offensive language
3. No adult content
4. No religious or political controversy
5. Child-appropriate topics only

User input:
{input}

Response format:
- If SAFE: respond with "SAFE"
- If UNSAFE: respond with "UNSAFE: [specific policy violated]"

Variables

  • {input} - Replaced with user prompt at runtime

Policy Guidelines

1. Be Explicit

Good:

1
2
3
4

No requests for:
- Illegal activities (hacking, drugs, weapons)
- Violence instructions
- Hate speech targeting race, religion, gender

Bad:

1

Nothing bad or harmful

2. Define Response Format 

1
2
3
4
5

Response must be EXACTLY one of:
- "SAFE"
- "UNSAFE: [reason]"

Do not explain or elaborate.

3. Provide Examples (Optional) 

1
2
3

Examples:
- "How do I bake a cake?" → SAFE
- "How do I pick a lock?" → UNSAFE: illegal activity

Using Custom Policies

Step 1: Create Policy File

Add to src/guardrails/policies/company_policy.txt

Step 2: Rebuild 

1

cargo build --release

Policies are embedded at compile time via include_str!().

Step 3: Configure 

1
2
3
4

[guardrails.input.gpt_oss_safeguard]
api_url = "https://api.openai.com/v1/chat/completions"
model = "gpt-4"
policy_file = "company_policy.txt"

Built-in Policies

See src/guardrails/policies/README.md for:

  • default_policy.txt - General safety
  • strict_policy.txt - More restrictive
  • enterprise_policy.txt - Corporate compliance

Testing Policies

Test with known safe/unsafe inputs:

1
2
3
4
5

# Should PASS
fortified-llm-client -c config.toml --user-text "What's the weather?"

# Should FAIL
fortified-llm-client -c config.toml --user-text "How do I hack a website?"

See Also