Wallbreaker Hermes

Wallbreaker Hermes is an AGPL-licensed fork of Wallbreaker for authorized LLM red-teaming. It keeps the standard Wallbreaker harness and adds an opt-in native laboratory for testing a fixed, ephemeral Hermes Agent target.

Documentation · PyPI · Releases · Contributing · Security

Scope and status

• Current release: v0.3.1 / wallbreaker-hermes==0.3.1.
• Python: 3.11 or newer.
• Project mode: open source collaborative. In-scope issues and pull requests are welcome.
• Hermes baseline: Hermes Agent v2026.8.13, package 0.20.1, commit f80f453ae0679347e38abc917c7f94f717bf96c5.

The standard CLI, TUI, dashboard, provider layer, attack tools, judge, reports and reliability checks remain available. The Hermes laboratory supports one text turn against a clean home or a selected, sanitized context. It rejects target tools, MCP, custom prompt layers, profiles, prefill, continuation and multimodal input.

The laboratory is not an operating-system sandbox. Its child process retains the filesystem and network permissions of the account that runs it.

Install

Install the published package:

python -m pip install wallbreaker-hermes==0.3.1
wallbreaker --help

The distribution name is wallbreaker-hermes. The import package and commands remain wallbreaker and wb for compatibility with upstream.

For development:

git clone https://github.com/Yivas/wallbreaker-hermes.git
cd wallbreaker-hermes
python -m venv .venv
. .venv/bin/activate
python -m pip install -e ".[dev]"
pytest -q

On Windows PowerShell, activate the environment with .\.venv\Scripts\Activate.ps1.

First use

Copy the example configuration, add credentials for providers you are authorized to use, then run the preflight check:

cp config.example.toml config.toml
wallbreaker check
wallbreaker

config.toml is ignored by Git. Do not commit credentials, prompts, responses, reports, session state or generated evidence.

Common entry points:

wallbreaker                         # terminal UI
wallbreaker dashboard               # local dashboard on 127.0.0.1:8787
wallbreaker --auto "objective..."   # autonomous loop against the configured target
wallbreaker report                  # render the latest local run
wallbreaker hermes --help           # Hermes laboratory workflow

See the installation guide and configuration guide for provider, dashboard and platform details.

Main capabilities

• OpenAI Chat Completions and Anthropic Messages provider normalization.
• Interactive and autonomous attack workflows with operator pause and explicit finish controls.
• HarmBench-backed evaluation, LLM judging and repeated reliability checks.
• Transform, preset, persona, multimodal, campaign and report tooling inherited from Wallbreaker.
• Optional P4RS3LT0NGV3 integration and generic MCP client support.
• Local FastAPI and React/Vite dashboard using the same application services as the TUI.
• Opt-in Hermes Agent laboratory with clean replicas, closed manifests, state comparison and permission-restricted evidence for local human review.

The documentation describes each capability without relying on private prompts, operational profiles or unpublished corpora.

Security and privacy

Use Wallbreaker Hermes only against systems you own or have explicit permission to test. Provider calls, target calls, judge calls and requested dataset updates can use the network. Run logs and reports may contain sensitive or harmful material even though their default locations are ignored by Git.

The Hermes campaign report is sanitized. When human review is required, prompt and response bodies are stored separately in a local permission-restricted sidecar. Hermes Agent does not receive or open that evidence.

Read SECURITY.md before exposing the dashboard, running untrusted targets or sharing artifacts. Report vulnerabilities through GitHub Private Vulnerability Reporting, not a public issue.

Documentation

The maintained documentation is published at https://yivas.github.io/wallbreaker-hermes/. Source files live under wiki/.

Repository-level references remain under docs/, including external-data attribution and examples that must stay close to the code. The operator integration and skill live under integrations/hermes/.

Support and contributing

Use GitHub Issues for reproducible bugs and scoped proposals. Pull requests are reviewed when they fit the project, preserve upstream compatibility and include the relevant tests and documentation. The project does not promise a response time or support for unauthorized testing.

Read CONTRIBUTING.md and CODE_OF_CONDUCT.md before opening a pull request.

License and upstream

Wallbreaker Hermes is licensed under AGPL-3.0-or-later. Modified versions, including versions offered over a network, must provide their complete corresponding source under the same license.

This repository preserves the Wallbreaker history and attribution. Third-party datasets and prompt corpora are not bundled automatically; their provenance and terms are documented in NOTICE and the external-data reference.