Open Source AI Security Tools

A robust Python library designed to enhance security and manage context when interacting with OpenAI's language models. Provides a protective layer for API calls, safeguarding against prompt injections, PII leaks, malicious URLs, and more. Supports multiple providers including OpenAI, Anthropic, and Cohere.

Key Features

• Secure API call wrapper for multiple LLM providers
• Input sanitization and output filtering
• Context management and validation
• Rate limiting and error handling
• Comprehensive logging and monitoring
• Intelligent response caching

GPU-accelerated logits processor implementing a shadow ban system to filter dangerous content during LLM text generation. Uses a vectorized Aho-Corasick automaton to detect banned patterns in O(1) time per token with zero inference overhead. Compatible with HuggingFace transformers and vLLM.

Key Features

• Vectorized Aho-Corasick engine with pre-computed danger mask
• Configurable shadow penalty levels (-5.0 to -20.0)
• Multi-level filtering with per-severity automata
• Automatic EOS forcing on complete pattern matches
• vLLM adapter for production deployments
• CLI tool for testing and validation

Bitmask-based LLM security firewall. A policy-driven LogitsProcessor that restricts what a language model can generate based on user permissions encoded as a capability bitmask. Built on resk-logits for GPU-accelerated pattern matching, with YAML policy configuration, hot-reload, and tool call prevention at the token level.

Key Features

• Capability bitmask permission system (up to 64 bits)
• Dual severity mode: hard block and bias penalty
• Tool call prevention at generation time, not post-hoc
• YAML policy with hot-reload and thread-safe cache
• Strict mode: forces EOS at first banned prefix
• No JWT handling -- receives raw mask from external auth

A comprehensive security toolkit for JavaScript and TypeScript applications using Large Language Models. Provides a wrapper around OpenAI-compatible APIs with built-in security features to protect your LLM integrations in web, Node.js, and frontend environments.

Key Features

• TypeScript-first security toolkit
• OpenAI API security wrapper
• Real-time threat detection
• Built-in rate limiting and caching
• Comprehensive type definitions

An open-source Python library that adds a robust security and management layer over the official Model Context Protocol (MCP) Python SDK. Enhanced security features, monitoring capabilities, and tools for managing MCP interactions.

Key Features

• Security wrapper for Model Context Protocol
• Enhanced monitoring and logging
• Protocol validation and sanitization
• Performance optimization tools
• Enterprise-grade error handling

The AI Agent Logger. Track every action your agents take with full context: function name, parameters, token probabilities, execution time, and results. Ship logs to Datadog, Prometheus, OpenTelemetry, webhooks, or local files. Designed for production observability with non-blocking async shipping.

Key Features

• Multi-destination logging: Datadog, Prometheus, OTel, webhooks, file
• Rich agent context: parameters, probabilities, duration, errors
• Async batch processing with configurable buffer
• Decorator-based and context manager APIs
• Framework agnostic: LangChain, CrewAI, AutoGen, custom

RESK Monitor

A monitoring solution for LLM deployments, offering comprehensive instrumentation and analysis capabilities with real-time error rate calculation and behavioral deviation detection.

In Development

ReskCrypt

An open-source project providing security and confidentiality in model discussions with end-to-end, state-of-the-art security level (ECDSA and AES).

Gnom

Advanced security tooling for enterprise environments. A comprehensive suite of security utilities designed for large-scale deployments and complex infrastructure management using Graph Neural Networks.

ReskLayer

Advanced prompt injection detection using ModernBERT with DiffTransformer attention mechanisms and Ettin's three-phase training recipe.