Prompt Injection Workbench

A research workbench for developing and testing attacks against large language models, with a focus on prompt injection vulnerabilities and defenses.

Key Features

• State Machine Design: Fine-grained control over agent execution for advanced attack scenarios
• SWE-bench Support: Benchmark agents on real-world code editing tasks from SWE-bench
• Hydra Configuration: Powerful experiment orchestration with parameter sweeps
• Extensible Architecture: Plugin system for custom agents, attacks, and environments
• Usage Limits: Built-in cost and resource controls
• Experiment Tracking: Automatic caching and result organization

Quick Start

Installation

Install the core package with desired optional features:

# Full installation (all features)
uv sync --all-extras

# Or install only what you need:
uv sync --extra agentdojo      # AgentDojo benchmark support
uv sync --extra swebench       # SWE-bench support
uv sync --extra docker         # Docker sandbox manager
uv sync --extra playwright     # Web automation environment

# Combine multiple extras
uv sync --extra agentdojo --extra docker

Available optional dependencies:

Extra	Description
agentdojo	AgentDojo dataset, environment, and attacks
swebench	SWE-bench dataset for code editing benchmarks
docker	Docker sandbox manager
playwright	Web automation environment

Set up environment variables:

cp .env.example .env  # Fill in API keys

Export default configuration:

# Export to ./config (default)
uv run prompt-siren config export

# Export to custom directory
uv run prompt-siren config export ./my_config

Run experiments:

# Run benign-only evaluation
uv run prompt-siren run benign +dataset=agentdojo-workspace

# Run with attack
uv run prompt-siren run attack +dataset=agentdojo-workspace +attack=agentdojo

# Run SWE-bench evaluation (requires Docker)
uv run prompt-siren run benign +dataset=swebench

# Run SWE-bench with specific instances
uv run prompt-siren run benign +dataset=swebench dataset.config.instance_ids='["django__django-11179"]'

# Run SWE-bench Lite (smaller benchmark)
uv run prompt-siren run benign +dataset=swebench dataset.config.dataset_name="SWE-bench/SWE-bench_Lite"

# Override parameters
uv run prompt-siren run benign +dataset=agentdojo-workspace agent.config.model=azure:gpt-5

# Parameter sweep (multirun)
uv run prompt-siren run benign --multirun +dataset=agentdojo-workspace agent.config.model=azure:gpt-5,azure:gpt-5-nano

# Validate configuration without running
uv run prompt-siren config validate +dataset=agentdojo-workspace

# Use config file with environment/attack included (no overrides needed)
uv run prompt-siren run attack --config-dir=./my_config

Tip: Environment and attack can be specified via CLI overrides or included directly in config files. See the Configuration Guide for details.

Analyzing Results

After running experiments, use the results command to aggregate and analyze results:

# View results with default settings (pass@1, grouped by all configs)
uv run prompt-siren results

# Specify custom output directory
uv run prompt-siren results --output-dir=./traces

# Group results by different dimensions
uv run prompt-siren results --group-by=model
uv run prompt-siren results --group-by=env
uv run prompt-siren results --group-by=agent
uv run prompt-siren results --group-by=attack

# Compute pass@k metrics (k>1)
uv run prompt-siren results --k=5
uv run prompt-siren results --k=10

# Compute multiple pass@k metrics simultaneously
uv run prompt-siren results --k=1 --k=5 --k=10

# Different output formats
uv run prompt-siren results --format=json
uv run prompt-siren results --format=csv

Understanding pass@k Metrics

• pass@1 (default): Averages scores across all runs for each task. A continuous metric showing average performance.
• pass@k (k>1): Binary success metric. A task "passes" if at least one of k runs achieves a perfect score (1.0). Uses statistical estimation when more than k samples are available.

Results Columns

The results table includes:

• Configuration columns: env_type, agent_type, attack_type, model, config_hash
• Metric columns: benign_pass@k, attack_pass@k - The pass@k scores
• Metadata columns:
  • n_tasks - Total number of tasks aggregated
  • avg_n_samples - Average number of runs per task
  • k - The k value (when computing multiple pass@k metrics)

Platform Requirements

• Python: 3.10+
• Package Manager for development: uv (for dependency management)
• Operating System: Linux or macOS (Windows not supported)
• Docker: Required for SWE-bench integration and some environments
  • Must be running and accessible
  • Base images should have /bin/bash available (Alpine images need bash package)

Documentation

• Configuration Guide - Hydra configuration and parameter overrides
• Usage Limits - Resource limits and cost controls
• Plugins - Adding custom agents, attacks, and environments

Development

# Lint and format
uv run ruff check --fix
uv run ruff format
uv run basedpyright

# Test
uv run pytest -v

License

Prompt Siren is licensed under an MIT License.