🚨 New open-source release: 𝚕𝚕𝚖-𝚜𝚊𝚗𝚒𝚝𝚒𝚣𝚎𝚛

LLM security is still lagging behind adoption. Prompt injection, hidden instructions in documents, and embedded agent directives are quickly becoming real attack vectors—not theoretical ones.

I’ve been working on a small but focused tool to address this gap:

👉 https://github.com/Warnes-Innovations/llm-sanitizer
👉 https://pypi.org/project/llm-sanitizer/

𝗪𝗵𝗮𝘁 𝗶𝘁 𝗱𝗼𝗲𝘀:

• Detects and classifies embedded LLM instructions in:

• documents
• source code
• web content
• Redacts or neutralizes those instructions before they reach your model
• Runs as:

• a Python package (importable directly in your pipeline)
• a CLI tool
• an MCP-compatible server

Conceptually, it sits as a 𝗽𝗿𝗲-𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗹𝗮𝘆𝗲𝗿—intercepting prompts before they ever hit your LLM.

This is an initial effort to address the growing need for 𝑑𝑒𝑓𝑒𝑛𝑠𝑒-𝑖𝑛-𝑑𝑒𝑝𝑡ℎ in LLM pipelines, especially given how easy it is to smuggle instructions via seemingly benign inputs (markdown, HTML, comments, etc.).

𝚕𝚕𝚖-𝚜𝚊𝚗𝚒𝚝𝚒𝚣𝚎𝚛 𝗲𝘅𝗽𝗹𝗶𝗰𝗶𝘁𝗹𝘆 𝘁𝗮𝗿𝗴𝗲𝘁𝘀 𝗲𝗺𝗯𝗲𝗱𝗱𝗲𝗱 𝗶𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻 𝗽𝗮𝘆𝗹𝗼𝗮𝗱𝘀 𝗮𝗻𝗱 𝗽𝗿𝗼𝗺𝗽𝘁 𝗶𝗻𝗷𝗲𝗰𝘁𝗶𝗼𝗻 𝘃𝗲𝗰𝘁𝗼𝗿𝘀 𝗮𝗰𝗿𝗼𝘀𝘀 𝗮𝗿𝘁𝗶𝗳𝗮𝗰𝘁 𝘁𝘆𝗽𝗲𝘀

─────────────

𝗪𝗵𝘆 𝗜 𝗯𝘂𝗶𝗹𝘁 𝘁𝗵𝗶𝘀

Most LLM applications today implicitly trust upstream content. That’s a risky assumption.

If your system ingests:

• scraped web pages
• PDFs / documents
• code repositories
• user-uploaded files

…then you already have a potential prompt injection surface.

─────────────

𝗪𝗵𝗮𝘁 𝗜’𝗱 𝗹𝗼𝘃𝗲 𝗳𝗲𝗲𝗱𝗯𝗮𝗰𝗸 𝗼𝗻

This is an early release (v0.1.0), and I’m actively iterating. I’d especially value input on:

• Detection strategy (regex vs. structural vs. model-assisted)
• False positive / false negative tradeoffs
• Integration patterns (pipelines, agents, MCP, etc.)
• Real-world attack examples I should include

─────────────

If you’re building in the LLM / agent space, I’d really appreciate you kicking the tires and sharing thoughts—good or bad.

Security here is still an open problem, and the ecosystem needs more experimentation.

─────────────

#LLM #AI #Security #PromptInjection #OpenSource #MLOps
