LLM-Filter-Probe: A Tool for Reverse-Engineering and Analyzing Input Censorship Mechanisms of Large Language Models

LLM-Filter-Probe is an open-source tool designed to reverse-engineer and analyze the input censorship mechanisms of large language model (LLM) API gateways, helping users quickly identify blocked sensitive keywords. It addresses the issue of opaque filtering rules on API platforms and is suitable for enterprise compliance teams, security researchers, developers, and other groups, enhancing the transparency and efficiency of compliance management.

As LLMs are widely applied across various industries, content security and compliance review have become key links in deployment. Major API service providers (such as OpenAI, Anthropic, NewAPI, OneAPI, etc.) implement strict keyword filtering at the gateway layer, but the rules exist in a black box form—users cannot know the specific triggering keywords and mechanisms, which brings troubles to compliance audits, security research, and developers. Against this background, LLM-Filter-Probe came into being.

LLM-Filter-Probe is a precise keyword detection and reverse-engineering tool that supports multi-platform APIs (NewAPI, OneAPI, etc.). Its core features include: 1. Precise keyword identification: Locating sensitive keywords through systematic testing; 2. Multi-platform support: Compatible with mainstream API gateways and extensible; 3. User-friendly interface: Simple and intuitive, ready to use out of the box, accessible even to non-technical personnel.

The tool's working principle consists of four steps: 1. Input sampling and mutation: Generate mutated test cases (word-by-word replacement, segment testing, etc.); 2. API interaction and response analysis: Send test cases and parse API responses (error codes/rejection messages); 3. Precise keyword localization: Compare blocked and passed cases, use algorithms like binary search to narrow down the scope; 4. Result output: Generate a structured report listing sensitive keywords and their contexts.

The tool's application scenarios include: 1. Enterprise compliance audits: Establish an understanding of censorship mechanisms to avoid false triggers; 2. Security research and red team testing: Evaluate the robustness of platform filtering; 3. Prompt engineering optimization: Improve API call success rates; 4. Cross-platform strategy comparison: Select API platforms suitable for business needs.

System requirements: Supports Windows 10+, macOS 10.14+, Linux (Python 3.6+), memory ≥4GB, disk space ≥200MB. Usage process: 1. Launch the application; 2. Input the text to be analyzed; 3. Click 'Analyze'; 4. View the list of sensitive keywords and their contexts.

Notes to consider: 1. Dynamic nature of censorship mechanisms: Rules may be updated, so regular testing is required; 2. Legal and ethical boundaries: Comply with laws and service terms, do not bypass maliciously; 3. Limitations of test coverage: Cannot cover 100% of all edge cases.

Future plans: Expand API type support, improve keyword analysis algorithms (introduce machine learning), and extend user guides. Conclusion: LLM-Filter-Probe enhances the transparency of LLM censorship mechanisms, facilitates compliance and security, and will play an important role in the field of AI security.