mhcoen

mhcoen / guardllm

Public

Hardening pipelines to protect LLMs from untrusted content

14
3
100% credibility
Found Feb 19, 2026 at 13 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

GuardLLM is a standalone Python library that adds fast local security layers to LLM apps processing untrusted content like web results, emails, and tool calls.

How It Works

1
đź’ˇ Discover GuardLLM

While building a smart helper that reads emails and web pages, you find a simple safety shield to protect against tricks in untrusted info.

2
🛡️ Add safety kit

Get the safety tools ready in seconds so your helper stays secure.

3
🧹 Clean first web result

Watch hidden tricks get scrubbed from a web page, leaving safe text ready for your helper.

4
🔧 Set simple rules

Tell your helper which actions like sending emails need extra okay.

5
âś… Check before acting

Before emailing or using tools, confirm everything is safe and approved.

🎉 Safe helper ready

Your smart assistant now handles risky info securely, fast, and without surprises.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 13 to 14 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is guardllm?

GuardLLM is a Python library that hardens LLM pipelines to protect against untrusted content like web results, emails, calendars, and MCP tool traffic. It sanitizes inputs, detects prompt injections, enforces tool policies, and tracks provenance—all locally in under 0.1ms, model-agnostic with no external APIs. Users get isolated content wrappers, action gates, outbound DLP, and audit logs via simple API calls like `guard.process_inbound()`.

Why is it gaining traction?

It composes fast heuristics (10,000x quicker than ML alternatives) with shared security context across ingress to outbound, scoring 85% F1 on injections and 100% on non-text controls—beating OpenAI, Anthropic, and Bedrock in benchmarks. Devs like the zero-latency local runs and easy integration for safe pipelines, unlike fragmented tools lacking lifecycle awareness.

Who should use this?

LLM engineers building agents with RAG/web search, MCP clients/servers, or email/calendar integrations; teams hardening github actions, linux/ubuntu setups, or windows/11 environments against untrusted content injections.

Verdict

Grab it for quick LLM hardening—tutorials and benchmarks make evaluation straightforward. 1.0% credibility score and 13 stars signal early maturity with solid tests/docs; prototype aggressively before production commit.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.