# NOTICE

superred-target-agentdojo

This package (all code under `src/agentdojo_target/`, including the original
`system_specification.md`) is original work of the superred project, released
under the MIT License. Copyright (c) 2026 Simon Sure.

## Upstream: AgentDojo

This package is a faithful port / adapter of the AgentDojo benchmark. It does
NOT vendor any AgentDojo source code or data: `agentdojo` is a pinned runtime
dependency (`agentdojo==0.1.35`) installed from PyPI, and all benchmark
environment data (the banking / workspace / slack / travel suite
`environment.yaml` and `injection_vectors.yaml` files) is loaded at run time
from the installed `agentdojo` package. Nothing from AgentDojo is redistributed
inside this wheel.

Because this module reproduces AgentDojo's tool set, task semantics, and
benchmark structure (pinned to AgentDojo v1.2.2; see `ASSUMPTIONS.md` for every
documented divergence), we credit the upstream authors:

- Project: AgentDojo (Debenedetti et al., "AgentDojo: A Dynamic Environment to
  Evaluate Attacks and Defenses for LLM Agents", NeurIPS 2024,
  arXiv:2406.13352)
- Repository: https://github.com/ethz-spylab/agentdojo
- License: MIT
- Copyright (c) 2024 Edoardo Debenedetti, Jie Zhang, Mislav Balunovic,
  Luca Beurer-Kellner, Marc Fischer, and Florian Tramèr

## Citation

When reporting results produced with this target, cite the AgentDojo paper
(Debenedetti et al., NeurIPS 2024).
