grith

Grith is a security proxy designed to enforce zero trust principles for AI agents that interact with operating systems and developer workflows. The platform addresses the risk of AI agents autonomously making security decisions, such as reading sensitive files, executing shell commands, or sending network requests, by intercepting every action before execution and verifying its safety. Grith is positioned as a solution to vulnerabilities like prompt injection, where malicious instructions can lead agents to exfiltrate credentials or perform unauthorized operations without alerting the user.

The tool operates by wrapping any CLI-based AI agent with a single command, requiring no modification to the agent itself. It performs OS-level syscall interception, capturing all file operations, network connections, and process spawns. Each intercepted action is evaluated by 18 independent security filters across three phases: static checks, pattern matching, and context analysis. Filters include path matching, secret scanning, taint tracking, behavioral profiling, and destination reputation, producing a composite score for each operation. Based on this score, actions are automatically allowed, queued for batch review in a quarantine digest, or denied and logged. Grith claims to allow 80-90% of actions automatically, with 5-15% queued for review and 1-5% denied.

Grith integrates with a wide range of AI coding agents and LLM providers, including Claude Code, Codex, Aider, Cline, Goose, Gemini CLI, Open Interpreter, as well as providers like Ollama, Anthropic, OpenAI, Google Gemini, and OpenRouter. It is designed to fit into existing developer workflows without requiring changes to models or providers. The platform provides detailed session summaries after each agent run, including tool call breakdowns, costs, and security outcomes, as well as structured audit trails in JSON format for compliance and analytics purposes. Grith supports export of security telemetry via webhook and JSON for integration with SIEM or SOAR systems, and is designed to aid compliance with frameworks such as SOC 2, NIST AI RMF, and HIPAA, though it is not itself certified.

For teams and enterprises, Grith offers features like encrypted team key synchronization and plans for SSO, RBAC, air-gapped deployment, and OpenTelemetry support. The tool is aimed at developers, security professionals, and organizations seeking to independently enforce security controls on AI agent actions and maintain auditable records of all operations.