# NOTICE

security-claim-agentharm (superred-claim-agentharm) is an original superred
SecurityClaim released under the MIT License, Copyright (c) 2026 Simon Sure.

It does NOT bundle any third-party source code or dataset rows. All AgentHarm
functionality (harmful tools, per-behavior grading functions, refusal and
semantic judges, prompt templates, and the dataset loader) is IMPORTED at
runtime from the `inspect-evals` and `inspect-ai` packages; the AgentHarm
dataset itself is DOWNLOADED at runtime from HuggingFace and is not
redistributed here. This NOTICE credits those upstreams and records the
runtime obligations they carry.

## This module's code

MIT, Copyright (c) 2026 Simon Sure. See `LICENSE`. The only bundled data file,
`src/agentharm_claim/data/benign_covers.json` (41 "benign cover" tasks), is
original work authored for superred and is covered by this same MIT license. It
is NOT derived from AgentHarm's benign split. The cover tasks reference some
AgentHarm entity identifiers (proper names, emails, a street address, X handles,
and tool names) purely for grader compatibility; these are non-copyrightable
facts and do not make the file a redistribution of the AgentHarm dataset.

## Inspect AI (runtime dependency)

- Source: https://github.com/UKGovernmentBEIS/inspect_ai
- License: MIT (SPDX: MIT), Copyright (c) 2024 UK AI Security Institute
- Used as: an unmodified pip dependency (`inspect-ai`). Tools, models, and
  dataset primitives are imported, not copied.

## Inspect Evals / AgentHarm implementation (runtime dependency)

- Source: https://github.com/UKGovernmentBEIS/inspect_evals
- License: MIT (SPDX: MIT), Copyright (c) 2024 UK AI Security Institute
- Used as: an unmodified pip dependency (`inspect-evals==0.13.1`). This module
  imports AgentHarm's harmful tool factories, grading functions, RefusalJudgeLLM
  and semantic judge, system/jailbreak prompt functions, and dataset loader
  from `inspect_evals.agentharm` verbatim (by import). `metrics.py` is an
  independent reimplementation of the arithmetic of
  `inspect_evals.agentharm.metric`.

## AgentHarm benchmark and dataset (Andriushchenko et al., 2024)

- Paper: "AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents",
  arXiv:2410.09024
- Dataset: https://huggingface.co/datasets/ai-safety-institute/AgentHarm
  (harmful `test_public` split, pinned revision
  e23b3fe60a0da9037314b88e5ee3a0c054970dad)
- Dataset license: MIT License WITH an additional field-of-use clause,
  Copyright (c) 2024 Gray Swan AI and UK AI Safety Institute. The additional
  clause states: "We prohibit using the dataset and benchmark for purposes
  besides improving the safety and security of AI systems."
- Redistribution: this package does NOT bundle the dataset; it is downloaded on
  first use from HuggingFace by upstream `inspect_evals`. Users obtain the
  dataset directly under the above license, and the field-of-use restriction
  applies to their use of it.

When reporting numbers produced with this module, cite Andriushchenko et al.
2024 (BibTeX in the README).
