Grith is live – security proxy for AI coding agents (grith.ai)

Grith, a security proxy for AI coding agents, has officially launched, providing OS-level protection for models like Claude Code and Codex on Linux. This tool enforces a robust boundary around AI processes, ensuring they request permission before executing actions, addressing the risks associated with auto-approval mechanisms that many users default to. Version 0.3.2 allows easy installation on Linux x86_64 and arm64 systems, and utilizes ptrace with seccomp-BPF pre-filters to evaluate security-related syscalls against a set of 18 filters, categorized into static checks, pattern matching, and contextual analysis, without involving probabilistic models in the decision-making process. The significance of Grith lies in its approach to security by placing the decision-making authority outside of the AI models themselves, promising reduced vulnerabilities to exploitative prompts and unauthorized actions. It establishes a scoring system with three verdicts (ALLOW, QUEUE, DENY) based on risks assessed during zero-trust evaluations. Grith addresses common escapes and provides a transparent, offline solution with comprehensive logging, all while being free for basic use with options for enhanced features in paid tiers. The developers encourage community feedback to refine the tool and expand support to other platforms like macOS and Windows in future releases.