# NOTICE

superred-claim-asb

This module is a port of Agent Security Bench (ASB) into the superred
red-teaming framework. Original code in this module (the superred adaptation:
task, factories, loaders, the adversarial success judge, and tests) is licensed
MIT (see LICENSE). It bundles benchmark data and reproduces evaluation logic
from ASB; the upstream MIT license text is preserved verbatim in
`src/asb_claim/data/ASB_LICENSE` (and also in `LICENSES/ASB-MIT.txt`).

## Upstream: Agent Security Bench (ASB)

- Project: Agent Security Bench (ASB)
- Repository: https://github.com/agiresearch/ASB (commit 1f561dcc)
- Paper: "Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and
  Defenses in LLM-based Agents", Hanrong Zhang, Jingyuan Huang, Kai Mei,
  Yifei Yao, Zhenting Wang, Chenlu Zhan, Hongwei Wang, Yongfeng Zhang,
  ICLR 2025.
- License: MIT
- Copyright (c) 2024 AGI Research

### Bundled data (vendored verbatim, MIT)

From the ASB repository, under `src/asb_claim/data/`:

- `agent_task.jsonl` (10 scenario agents + benign tasks)
- `all_attack_tools.jsonl` (400 attacker tools)
- `all_attack_tools_aggressive.jsonl` (200 aggressive)
- `all_attack_tools_non_aggressive.jsonl` (200 non-aggressive)
- `all_normal_tools.jsonl` (20 normal tools, Expected Achievements strings)
- `ASB_LICENSE` (verbatim upstream MIT license)

### Ported evaluation logic (MIT)

The attack-success / utility (`check_original_success`) / refusal
(`judge_response`) / `memory_found` predicates are ported from ASB
`main_attacker.py`, and the refusal-judge system prompt
(`ASB_REFUSAL_SYSTEM_PROMPT` in `src/asb_claim/judge.py`) is
reproduced verbatim. Deviations are documented in this module's
`ASSUMPTIONS.md`.

The adversarial success judge (`AsbAdversarialJudge` /
`ASB_ADVERSARIAL_SYSTEM_PROMPT`), the deterministic tool-invocation success
predicate, the memory trust-boundary tagging, and the derived convenience
sub-scores are original to this module and have no upstream analogue.

## Citation

When reporting numbers produced with this module, cite Zhang et al., ICLR 2025
(Agent Security Bench).
