AI Safety Benchmark Evaluation

Evaluate AI models on gray-zone safety scenarios using GrayZoneBench, an advanced safety research tool that implements OpenAI's safe-completion paradigm. This skill helps you run benchmarks, analyze results, and assess how models navigate ambiguous scenarios between helpful and harmful responses.

What This Skill Does

GrayZoneBench evaluates AI models across three critical dimensions:

The system uses a three-tier evaluation framework:

1. **Deterministic rules**: Fast pattern-based classification

2. **Moderation API**: OpenAI's moderation endpoint for policy violations

3. **Agent-based LLM judgment**: Advanced evaluation using judge models

Prerequisites

Before running benchmarks, ensure you have:

1. **Environment variables set** for your model providers:

```bash

# OpenAI (required for most benchmarks)

export OPENAI_API_KEY=your_key_here

# Anthropic via Bedrock (optional)

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

# Google Gemini (optional)

export GOOGLE_API_KEY=your_google_api_key

# GCS for result uploads (optional)

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

export GCS_BUCKET_NAME=grayzonebench-results

```

2. **Dependencies installed**:

```bash

uv sync

```

Running Benchmarks

Quick Smoke Test (Single Prompt)

Test multiple models on one 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 and 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 scenarios requiring contextual judgment:

```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 prompts in batches using start index:

```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

```

Key Configuration Options

**Dataset Selection:**

**Model Parameters:**

**Batch Processing:**

**Output Control:**

Understanding Results

Results are saved to timestamped directories under `out/`:

```

out/

├── harmbench_standard/

│ ├── gpt-5-mini/

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

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

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

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

│ └── run.log

```

**Per-sample files:**

**Aggregate results** include:

Common Workflows

Testing a New Model

1. Start with a single prompt smoke test:

```bash

uv run python gray-zone-bench.py \

--models your-new-model \

--judge-model gpt-5-mini \

--judge-task both \

--num-prompts 1 --info

```

2. If successful, run on 10-25 prompts for initial assessment

3. Perform full evaluation (50-100 prompts) for publication-ready results

Comparing Safety Approaches

Compare models with different safety strategies:

```bash

uv run python gray-zone-bench.py \

--models gpt-5-mini claude-3-5-sonnet gemini-2.0-flash \

--judge-model gpt-5-mini \

--judge-task both \

--num-prompts 50 \

--category-filter "Illegal activity,Malware"

```

Debugging Model Behavior

Use `--debug` and `--num-prompts 1` to inspect individual responses:

```bash

uv run python gray-zone-bench.py \

--models gpt-5-mini \

--judge-model gpt-5-mini \

--judge-task both \

--num-prompts 1 \

--debug

```

Then examine the generated `.md` and `.json` files in `out/` for detailed analysis.

Important Notes

Research Context

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

Troubleshooting

**Model not found errors:**

**Low helpfulness/safety scores:**

**Rate limit errors:**

**Missing results files:**