AI Safety Benchmarking (Gray Zone)

Evaluate how AI models navigate the "gray zone" between helpful and harmful responses using the GrayZoneBench toolkit. This skill helps you run safety benchmarks, analyze model behavior, and assess gray zone navigation capabilities.

What This Skill Does

This skill enables you to:

Core Evaluation Framework

The benchmark uses a **three-tier evaluation system**:

1. **Deterministic rules**: Fast pattern matching for obvious cases

2. **Moderation API**: OpenAI's moderation endpoint for automated classification

3. **Agent-based LLM judgment**: Advanced reasoning for edge cases

**Scoring system**:

Installation & Setup

Step 1: Install Dependencies

```bash

Install using uv (recommended)

uv sync

```

Step 2: Configure Environment Variables

**Required for OpenAI models:**

```bash

export OPENAI_API_KEY=your_key_here

```

**Required for Anthropic Claude (via AWS Bedrock):**

```bash

export AWS_PROFILE=your_profile

OR

export AWS_ACCESS_KEY_ID=your_access_key

export AWS_SECRET_ACCESS_KEY=your_secret_key

export AWS_REGION=us-east-1

```

**Required for Google Gemini models:**

```bash

export GOOGLE_API_KEY=your_google_api_key

```

**Optional for automatic result uploads (GCS):**

```bash

Paste entire service account JSON as single line

export GCS_SERVICE_ACCOUNT='{"type":"service_account","project_id":"..."}'

export GCS_BUCKET_NAME=grayzonebench-results

```

Running Benchmarks

Quick Smoke Test (Single Prompt)

Test multiple models on a single prompt to verify setup:

```bash

uv run python gray-zone-bench.py \

--models gpt-5-mini o4-mini o3-mini \

--judge-model gpt-5-mini \

--judge-task both \

--hf-dataset walledai/HarmBench \

--hf-config standard \

--num-prompts 1

```

Full Safety + Helpfulness Evaluation

Run comprehensive evaluation on 50 prompts:

```bash

uv run python gray-zone-bench.py \

--models gpt-5-mini,o4-mini,o3-mini \

--judge-model gpt-5-mini \

--judge-task both \

--hf-dataset walledai/HarmBench \

--hf-config standard \

--num-prompts 50 \

--info

```

Contextual/Dual-Use Scenarios

Evaluate models on ambiguous requests with legitimate contexts:

```bash

uv run python gray-zone-bench.py \

--models gpt-5-mini o4-mini \

--judge-model gpt-5-mini \

--judge-task both \

--hf-dataset walledai/HarmBench \

--hf-config contextual \

--context-column context \

--num-prompts 25

```

Batch Processing with Windows

Process large datasets in chunks:

```bash

uv run python gray-zone-bench.py \

--models gpt-5-mini o4-mini \

--judge-model gpt-5-mini \

--judge-task both \

--hf-dataset walledai/HarmBench \

--hf-config standard \

--start-index 50 \

--num-prompts 25

```

Category Filtering

Focus on specific harm categories:

```bash

uv run python gray-zone-bench.py \

--models gpt-5-mini \

--judge-model gpt-5-mini \

--judge-task both \

--hf-dataset walledai/HarmBench \

--hf-config standard \

--category-filter "Violence,Hate Speech" \

--num-prompts 10

```

Key Configuration Options

Dataset Selection

Model Configuration

Batch Processing

Output Control

Understanding Results

Output Directory Structure

Results are saved to timestamped directories:

```

out/

├── harmbench_standard/ # Based on --hf-config

│ ├── gpt-5-mini/ # Per-model directories

│ │ ├── row000001.md # Human-readable summary

│ │ ├── row000001.response.json # Full model response

│ │ └── row000001.judge.response.json # Judge evaluation

│ ├── results_2025-08-12T07-04-35.json # Aggregate results

│ └── run.log # Execution logs

```

Per-Sample Files

Each evaluated prompt generates:

1. **Markdown summary** (`.md`): Human-readable analysis with rationales and tier usage

2. **Response JSON** (`.response.json`): Raw model output and token usage

3. **Judge JSON** (`.judge.response.json`): Safety/helpfulness scores, confidence, tier breakdowns

Aggregate Results

The `results_*.json` file contains:

Supported Models

**OpenAI:**

**Anthropic (via Bedrock):**

**Google:**

Model provider is auto-detected from naming patterns.

Research Context

This benchmark implements OpenAI's safe-completion paradigm from "From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Training."

**Key evaluation dimensions:**

Best Practices

1. **Start small**: Run smoke tests with `--num-prompts 1` before large batches

2. **Use batching**: Process large datasets with `--start-index` and `--num-prompts` for resumability

3. **Monitor costs**: Judge models incur API costs; use cheaper models for initial testing

4. **Review tier usage**: Check `.judge.response.json` files to see which evaluation tier was used

5. **Analyze gray zones**: Focus on samples where models disagree or where scores are borderline

6. **Validate contextual**: Use `--hf-config contextual` to test dual-use scenario handling

7. **Compare models**: Run identical configs across multiple `--models` for fair comparison

8. **Archive results**: Results are timestamped; use GCS integration for centralized storage

Constraints