The first bi-directional LLM prompt firewall.

PromptLeakFence is a lightweight, transparent security proxy designed to prevent prompt injection attacks and unauthorized sensitive information disclosure (PII, credentials, etc.) in AI systems. It sits between your application and your LLM provider, scanning every message for potential leaks.

• 🔍 Real-time Interception: Acts as a transparent proxy for LLM APIs.
• 🔒 Bi-directional Filtering: Scans both incoming prompts and outgoing assistant responses.
• 🛡️ Policy-based Redaction: Automatically redacts sensitive patterns (passwords, API keys, tokens) using customizable rules.
• 🚀 Multi-Provider Support: Pre-configured for Ollama, llama.cpp, LMStudio, Gemini, Claude, and more.
• 📊 Audit Dashboard: Specialized dashboard to visualize and analyze blocked attempts.
• 🛠️ Lightweight & Fast: Built with Go for high performance and low latency.

• Go 1.25+
• Make

make build

This will generate the plf binary in the bin/ directory.

The proxy requires a target LLM service to forward clean traffic to.

# Example: Running PLF in front of Ollama
sudo ./bin/plf proxy --target ollama --listen 8080 --verbose

Supported Targets:

• ollama (default port: 11434)
• llama.cpp (8080)
• lmstudio (1234)
• oobabooga (7860)
• openwebui (3000)
• copilot (5000)
• gemini (8080)
• claude (8080)

Visualize blocked leak attempts:

./bin/plf audit --port 9090

Then visit http://127.0.0.1:9090.

Security rules are defined in configs/policy.yaml. You can customize the exclude list to add patterns that should be redacted.

policy:
  exclude:
    - "token"
    - "password"
    - "api_key"
    # ... add your custom sensitive keywords here

Every match found in a prompt or response will be replaced by [INTERNAL_PROMPT_REDACTED].

• main.go: Entry point for the CLI.
• cmd/: Command implementations (proxy, audit, root).
• cmd/utils/: Core logic for network redirection, policy parsing and launch helpers.
• configs/: Default security policies.

Contributions are welcome! Please feel free to submit a Pull Request.

See LICENSE file for details.