data-privacy-stack/presidio
Presidio is an open-source framework for detecting, redacting, masking, and anonymizing sensitive data (PII) across text, images, and structured data, leveraging NLP and customizable pipelines.
Awesome Infra for AI › AI Safety & Guardrails
DobermanCore/Doberman-Core
Doberman is a security layer for AI coding agents that intercepts every tool call, file write, shell command, or MCP tool invocation, before it executes, rather than only advising the model afterward. It attaches through a transparent MCP proxy or a host hook (native hooks for Claude Code, an experimental PreToolUse hook for Codex, a plugin adapter for OpenClaw, and experimental native hooks for Cursor), normalizes the action, runs it through a configurable risk engine, and returns one of three verdicts: PASS lets routine work through unchanged, AUTH pauses for human approval, with a short-lived re-prompt window for a repeated identical action, and BLOCK stops dangerous actions outright. The project states two enforcement guarantees: it fails closed, so any error, timeout, or unhandled case denies the action rather than allowing it, and an unanswered approval prompt resolves to denial after a fixed deadline; and it is raise-only, meaning automated adaptive learning can tighten policy but any permanent loosening requires an explicit, human-approved, second-factor-gated action. Policy is expressed as a repo-committed YAML file that can be reviewed like code, and users can register their own guardrail plugins to add custom rules or audit sinks. A benchmark compares the attack-block rate against the false-positive rate, and a parity matrix documents which protections are proven, per host, by a linked CI test. It ships as a pip-installable CLI with setup, uninstall, and diagnostic commands, plus an optional local dashboard for reviewing blocked actions and pending approvals. It targets teams and individual developers running autonomous or semi-autonomous coding agents who want a policy-enforced execution boundary rather than relying solely on the model's own judgment.
https://github.com/DobermanCore/Doberman-Core
Also in AI Safety & Guardrails
NVIDIA-NeMo/Guardrails
NVIDIA NeMo Guardrails is an open-source toolkit for adding programmable guardrails to LLM-based conversational applications, focusing on safety, security, and controlled dialog.
superagent-ai/superagent
Superagent is an open-source SDK providing safety features for AI applications, including prompt injection detection, PII redaction, repository scanning for threats, and red teaming capabilities fo...
Tencent/AI-Infra-Guard
AI-Infra-Guard is a full-stack AI red teaming platform providing comprehensive security analysis, vulnerability scanning, and jailbreak evaluation for AI ecosystems and LLMs.
microsoft/agent-governance-toolkit
AI Agent Governance Toolkit (AGT) provides policy enforcement, identity management, execution sandboxing, and reliability engineering to secure autonomous AI agents in production.
FailproofAI/failproofai
Observability and policy enforcement for AI agent harnesses, hooking twelve coding and chat harnesses to record every run and block dangerous tool calls before they execute.
protectai/llm-guard
LLM Guard is a comprehensive open-source security toolkit designed to fortify Large Language Model (LLM) interactions by providing robust sanitization, malicious content detection, data leakage pre...
lennney/stop-that-shit
Multi-platform hook and skill guard for AI coding agents that blocks unrequested work such as generated hashes, checksums and task-scope expansion at the harness hook boundary.