amazon-science/PurpleAudit

amazon-science/PurpleAudit

Language: Python

License: NOASSERTION

Stars: 0

Forks: 0

Open issues: 0

Created: 2026-10-22T17:48:57Z

Pushed: 2026-10-22T18:16:38Z

Default branch: main

Fork: no

Archived: yes

README:

PurpleAudit

Code for PurpleAudit: A Co-Evolutionary Auditing Framework for Task Hijacking in Multi-Agent Systems.

PurpleAudit audits a multi-agent system (MAS) for task hijacking by indirect prompt injection. It runs your MAS on your tasks, plants an attacker's request in the output of the first tool your agents call, and measures the attack success rate (ASR) and utility (tasks still solved), with and without defenses. A red-team / blue-team co-evolution then finds stronger attacks against your MAS and a defense prompt that counters them. MASs built with LangGraph, CrewAI or AG2 are audited without changing their code.

Install

cd PurpleAudit
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[langgraph,crewai]"
cp .env.example .env # then set OPENROUTER_API_KEY (or use AWS Bedrock, see docs/guide.md)

Quick start: audit an example

[examples/default/](examples/default/) holds three small MASs: langgraph_support (LangGraph), crewai_travel (CrewAI) and ag2_helpdesk (AG2).

MAS=langgraph_support # or crewai_travel, ag2_helpdesk
MODEL=us.anthropic.claude-sonnet-4-5-20250929-v1:0

# 1. Benign runs, then 2 of the paper's evolved attacks (drop --attacks to run all 28 attacks)
python -m purpleaudit.framework $MAS --folder examples --benign --attack --attacks evolved_mas_specialized_agent_access_credential_injection,evolved_synchronization_lock_injection

# 2. The same attacks under a defense
# (spotlighting, prompt_sandwiching, tool_filtering, naive_safeguard, tool_output_extraction)
python -m purpleaudit.framework $MAS --folder examples --attack --attacks evolved_mas_specialized_agent_access_credential_injection,evolved_synchronization_lock_injection --defense-id tool_output_extraction

# 3. ASR and utility, without and with the defense
python -m purpleaudit.evaluate per-env --folder examples --model $MODEL
python -m purpleaudit.evaluate per-env --folder examples --model $MODEL --defense-id tool_output_extraction

# 4. Co-evolution: new attacks against the MAS and a defense prompt for it, then their scores
python -m purpleaudit.evolve --folder examples --env $MAS --iterations 2 --num-attacks 2 --evolution-name demo --attacks evolved_mas_specialized_agent_access_credential_injection,evolved_synchronization_lock_injection
python -m purpleaudit.evaluate per-env --folder examples --model $MODEL --iteration 1 --evolution-name demo
python -m purpleaudit.evaluate per-env --folder examples --model $MODEL --iteration 1 --evolution-name demo --defense-id 0

Transcripts and verdicts land in examples///, evolved attacks and defense prompts in examples/_evolution/demo/. Each example also runs on its own, e.g. python examples/default/langgraph_support/support_app.py "What is the status of order ORD-1001?".

Audit your own MAS

1. Create three files (see the [examples](examples/default/)):

audits/default/my_mas/
├── environment.json {"entrypoint": "audit:run"}
├── prompts.json {"task_1": {"prompt": "...", "ground_truth": "..."}, ...}
└── audit.py

2. `audit.py` runs your MAS, unchanged, on one task:

from my_app import answer # your code, next to audit.py or installed

def run(task, ctx):
return answer(task, ctx.chat_model())

While run executes, PurpleAudit instruments LangGraph / LangChain, CrewAI and AG2: the attack lands in the first tool output, defenses are added to tool outputs and system prompts, and the conversation is recorded. Passing the model under audit is optional (it lets --model pick your MAS's model): ctx.chat_model() for LangChain / LangGraph, ctx.crewai_llm() for CrewAI, ctx.ag2_llm_config() for AG2. For code outside these frameworks, see [hooks](docs/guide.md#hooks).

3. Write tasks in prompts.json: at least two tasks that make your MAS call a tool, each with a short, distinctive ground_truth that a correct answer contains.

4. Run the quick-start commands with MAS=my_mas and --folder audits.

Magentic-One on AssistantBench / GAIA

The paper's real-world MAS: Microsoft's Magentic-One, a web-browsing and coding team, on the AssistantBench and GAIA benchmarks. The tasks come from HuggingFace; GAIA is gated, so accept its terms on its page and log in with huggingface-cli login first. AssistantBench has 33 tasks (about $33 per audit in the paper). Magentic-One browses the web and runs the code it writes on your machine, so prefer a container or VM.

pip install -e ".[benchmarks]" && playwright install chromium
python -m purpleaudit.framework assistant_bench --benign --attack --attacks evolved_mas_specialized_agent_access_credential_injection,evolved_synchronization_lock_injection
python -m purpleaudit.evaluate per-env --model $MODEL --env assistant_bench

More

• [docs/guide.md](docs/guide.md): how audits work, options, outputs, spec-based MASs, tests.
• [docs/paper.md](docs/paper.md): reproducing the paper's experiments.

Release notice

This code is being released solely for academic and scientific reproducibility purposes, in support of the methods and findings described in the associated publication. Pull requests are not being accepted in order to maintain the code exactly as it was used in the paper.

Security

See [CONTRIBUTING](CONTRIBUTING.md#security-issue-notifications) for more information.

License

This project is licensed under the CC-BY-NC-4.0 License. See [LICENSE](LICENSE).