Metadata-Version: 2.4
Name: superred-target-asb
Version: 0.1.0
Summary: Agent Security Bench (ASB) composite agent target for superred: runs ASB's real plan-then-execute agent loop (vendored, pinned to upstream 1f561dcc) with event-based injection at the four ASB surfaces (DPI/OPI/PoT/MP), routed through a litellm proxy.
Project-URL: Homepage, https://superred.simonsure.com
Project-URL: Documentation, https://superred.simonsure.com
Project-URL: Repository, https://github.com/RoldSI/superred-modules
Project-URL: Issues, https://github.com/RoldSI/superred-modules/issues
Author-email: Simon Sure <info@simonsure.com>
License-Expression: MIT
License-File: LICENSE
License-File: LICENSES/ASB-MIT.txt
License-File: NOTICE
Keywords: agent,agent-security-bench,ai-safety,llm,red-teaming,security,superred
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Topic :: Security
Requires-Python: <3.14,>=3.11
Requires-Dist: jsonlines>=4.0
Requires-Dist: litellm>=1.40
Requires-Dist: numpy>=1.26
Requires-Dist: openai>=1.20
Requires-Dist: pandas>=2.0
Requires-Dist: python-dotenv>=1.0
Requires-Dist: superred>=0.1.0
Provides-Extra: dev
Requires-Dist: pytest-asyncio>=0.23; extra == 'dev'
Requires-Dist: pytest-cov>=6.0; extra == 'dev'
Requires-Dist: pytest>=8.0; extra == 'dev'
Description-Content-Type: text/markdown

# superred-target-asb

Agent Security Bench (ASB) agent as a [superred](https://superred.simonsure.com)
`Target`.

This module runs ASB's real, vendored plan-then-execute agent loop (pinned to
upstream commit `1f561dcc`) against a litellm proxy. It is a bare runtime that
exposes the four ASB injection surfaces (DPI / OPI / PoT / MP) as superred
Controllables over a trust-boundary forest (roots: user, system, tools,
memory), restores ASB's durable memory store, and performs no injection by
default. The benchmark-specific data, grading, and tasks live in the paired
`superred-claim-asb` package.

The importable package name is `asb_target`:

```python
from asb_target import AsbTarget
```

## Install

```bash
pip install superred-target-asb
```

## Credits / upstream

This target vendors and runs the real agent loop from **Agent Security Bench
(ASB)** by AGI Research (ICLR 2025, arXiv:2410.02644), pinned to upstream
commit `1f561dcc`.

- ASB source (`pyopenagi/`, `aios/`) is redistributed under its original
  **MIT License, Copyright (c) 2024 AGI Research**
  (https://github.com/agiresearch/ASB), preserved at
  `src/asb_target/_asb_vendor/LICENSE`.
- The bundled tool-definition data (`data/all_normal_tools.jsonl`) and the
  example-agent `config.json` files are ASB data under the same MIT license.

Everything else in this module (the superred Target, the event-based injection
model, the litellm-proxy transport, the durable cosine memory store, and the
trust-boundary forest) is original work, MIT-licensed, Copyright (c) 2026
Simon Sure. See `LICENSE` and `NOTICE`. Cite ASB when reporting results.
