AgentDyn: Are Your Agent Security Defenses Deployable in Real-World Dynamic Environments?

May 19, 2026 ยท View on GitHub

Hao Li, Ruoyao Wen, Shanghao Shi, Ning Zhang, Yevgeniy Vorobeychik, Chaowei Xiao.

The official implementation of the paper "AgentDyn: Are Your Agent Security Defenses Deployable in Real-World Dynamic Environments?".

AgentDyn is a dynamic, open-ended agent security benchmark featuring 60 challenging open-ended user tasks and 560 injection test cases across the Shopping, GitHub, and Daily Life scenarios. It is built on top of the AgentDojo framework. A huge thanks to the AgentDojo team for their admirable contribution to the community!

Quickstart

pip install -e .

Running the benchmark

For adaptability, we support an evaluation script same as AgentDojo's. Documentation on how to use the script can be obtained with the --help flag.

For example, to run the shopping suite , with gpt-4o-2024-08-06 as the LLM, the tool filter as a defense, and the attack with important_instructions, run the following command:

python -m agentdojo.scripts.benchmark -s shopping \
    --model GPT_4O_2024_08_06 \
    --defense tool_filter --attack important_instructions

To run with external defenses integrated in this repo layout, you can directly use:

Before running, please export your API key, through:

  1. OpenAI Model: export OPENAI_API_KEY=XXX
  2. Google Model: export GOOGLE_API_KEY=XXX
  3. Open-sourced Models (Qwen, LlaMA, other supported models): export OPENROUTER_API_KEY=XXX

Supported settings

Available Suites:

AgentDyn supports shopping,github, and dailylife suites, as well as the original four suites from AgentDojo (banking,slack, travel and workspace).

Available Models:

We evaluate the following models in our paper: GPT_4O_MINI_2024_07_18, GPT_4O_2024_08_06, GEMINI_2_5_FLASH, GEMINI_2_5_PRO, LLAMA_3_3_70B, QWEN3_235B, GPT_5_1_2025_11_13, GPT_5_MINI_2025_08_07. Other models supported by AgentDojo are also compatible.

Available Defenses:

In addition to the original defenses in AgentDojo, we provide support for PIGuard and PromptGuard2. We also support directly invoking external defenses in this workspace: CaMeL, Progent, DRIFT,.

The complete list of defenses supported in our paper includes: repeat_user_prompt, spotlighting_with_delimiting, tool_filter, transformers_pi_detector, piguard_detector, prompt_guard_2_detector, camel, progent, drift.

Inspect Results

To review the results reported in our paper, please refer to the log files in the (runs/).

References

If you find this work useful in your research or applications, we appreciate that if you can kindly cite:

@articles{AgentDyn,
  title={AgentDyn: Are Your Agent Security Defenses Deployable in Real-World Dynamic Environments?},
  author={Hao Li and Ruoyao Wen and Shanghao Shi and Ning Zhang and Yevgeniy Vorobeychik and Chaowei Xiao},
  journal = {arXiv},
  eprint = {2602.03117},
  year={2026}
}