semfire 0.2.7

SemFire (Semantic Firewall): detect advanced AI deception, including in-context scheming and multi-turn manipulative attacks.

SemFire

AI Deception Detection Toolkit

SemFire (Semantic Firewall) is an open-source toolkit for detecting advanced AI deception, with a primary focus on "in-context scheming" and multi-turn manipulative attacks. This project aims to develop tools to identify and mitigate vulnerabilities like the "Echo Chamber" and "Crescendo" attacks, where AI models are subtly guided towards undesirable behavior through conversational context.

Project Vision: A Toolkit for AI Deception Detection

History

SemFire aims to be a versatile, open-source toolkit providing:

• A Python library for direct integration into applications and research.
• A Command Line Interface (CLI) for quick analysis and scripting.
• A REST API service (via FastAPI) for broader accessibility and enterprise use cases.
• Core components that can be integrated into broader semantic-firewall-like systems to monitor and analyze AI interactions in real-time.

Features

• Rule-based detector (EchoChamberDetector) for identifying cues related to "in-context scheming," context poisoning, semantic steering, and other multi-turn manipulative attack patterns (e.g., "Echo Chamber", "Crescendo").
• Analyzes both current text input and conversation history to detect evolving deceptive narratives.
• Heuristic-based detector (HeuristicDetector) for signals like text complexity and keyword usage.
• ML-based classifiers to enhance detection of complex scheming behaviors over extended dialogues (Future Work).

API Instructions forthcoming.

• Python API for programmatic access.
• REST service (FastAPI) for network-based access.

Installation

The project can be installed from PyPI:

pip install semfire

• Quickstart :/docs/quickstart.md
• Containerized CLI : /docs/docker-cli.md
• Usage : /docs/usage.md
• LLM Providers for ai-as-judge features : /docs/providers.md

Terminal Demos (GIFs)

Examples

The following terminal demo GIFs are available under assets/demos/asciinema/:

Quick Start

Individual Detectors

Python API

Complete Workflow

API: Health/Ready/Zip Analyze

Contributing

Contributions are welcome! Please open an issue or submit a pull request.