Agentic Misalignment Research

What it does

• Research framework that probes frontier models with fictional scenarios — including blackmail and information leakage — to study the potential for agentic misalignment behavior
• Three-step pipeline: scripts/generate_prompts.py → run_experiments.py → classify_results.py, labeling each response with classification.harmful_behavior true/false
• Combinatorial experiment design: scenarios (murder, blackmail, leaking) × goal types (explicit, none, latent, ambiguous, swap) × urgency types (replacement, restriction, none)
• Runs 40+ models across 5 providers (Anthropic, OpenAI, Google, Together, OpenRouter) with provider-parallel execution (a 3–4x speedup), rate-limit handling, and auto-resume on interruption

When to use it

A safety reference for teams shipping production agents: reproduce Anthropic's misalignment probes under your own configs (configs/example_experiment_config.yaml), or extend the scenario matrix to your own fictional scenarios.

Getting started

git clone https://github.com/anthropic-experimental/agentic-misalignment.git
cd agentic-misalignment
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env  # add provider API keys

Then generate prompts and run experiments with --config configs/example_experiment_config.yaml (scripts/generate_prompts.py, scripts/run_experiments.py), and classify with scripts/classify_results.py --results-dir results/[experiment_id].

Things to know

Lives under the anthropic-experimental org. Requires API keys from multiple providers (Anthropic/OpenAI/Google/Together/OpenRouter); on rate-limit errors, lower provider concurrency in the config.