GPT OSS Safeguard

Custom policy-based validation using GPT-4.

Overview

GPT OSS Safeguard uses GPT-4 (or similar models) to validate inputs against custom policies defined in text files.

Advantages:

  • Fully customizable policies
  • Natural language policy definitions
  • High accuracy (GPT-4 level)

Disadvantages:

  • Slower (2-5 seconds)
  • More expensive (GPT-4 API costs)

Configuration 

1
2
3
4
5
6
7
8
9
10

[guardrails.input]
type = "gpt_oss_safeguard"

[guardrails.input.gpt_oss_safeguard]
api_url = "https://api.openai.com/v1/chat/completions"
model = "gpt-4"
api_key_name = "OPENAI_API_KEY"
policy_file = "policies/my_policy.txt"  # Custom policy
max_tokens = 1000
timeout_secs = 120

Policy Files

Location: src/guardrails/policies/ (embedded at compile time)

Built-in Policies

  1. default_policy.txt - General safety (violence, hate speech, illegal content)
  2. strict_policy.txt - Stricter rules (includes controversial topics)
  3. enterprise_policy.txt - Corporate compliance (data privacy, IP protection)

Custom Policy Example

policies/my_policy.txt:

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

You are a safety validator. Evaluate the following user input and determine if it violates any of these policies:

1. No requests for illegal activities
2. No hate speech or discrimination
3. No personal data collection without consent
4. No medical advice (we are not licensed)
5. No financial advice (we are not qualified)

If the input violates any policy, respond with "UNSAFE: [policy number]".
If the input is safe, respond with "SAFE".

User input to evaluate:
{input}

Variables:

  • {input} - User prompt to validate

Usage 

1
2
3

export OPENAI_API_KEY=sk-...
fortified-llm-client -c config.toml --user-text "Give me medical advice"
# Result: ValidationError (UNSAFE: policy 4 - medical advice)

Creating Custom Policies

Step 1: Write Policy File

Create src/guardrails/policies/my_custom_policy.txt:

1
2
3
4
5
6
7
8

Evaluate this input against our company policies:
- No competitor mentions
- No profanity
- Professional tone required

Input: {input}

Respond: SAFE or UNSAFE: [reason]

Step 2: Rebuild 

1

cargo build --release

Policy is embedded at compile time.

Step 3: Configure 

1
2

[guardrails.input.gpt_oss_safeguard]
policy_file = "my_custom_policy.txt"  # Just the filename

Best Practices

  1. Be specific in policies - Clear rules get better results
  2. Test with edge cases - Validate policy catches violations
  3. Set appropriate timeouts - GPT-4 can be slow (120s recommended)
  4. Monitor costs - Each validation is a GPT-4 API call

See Also