Show HN: AgentTrust – Zero-trust security and human approvals for AI agents (github.com)

AgentTrust has introduced a robust security framework designed specifically for autonomous AI agents, addressing the critical risks associated with unrestricted AI side effects such as hallucinations and infinite loops. This solution implements a zero-trust security model that ensures no AI agent can execute actions or access databases without explicit permission, utilizing features like cascading subagent delegation, human-in-the-loop approvals, and dynamic semantic blast radius scoring. The significance of this innovation lies in its ability to safeguard enterprises deploying AI agents from financial and data-related disasters while complying with security regulations. Key technical components of AgentTrust include a transparent egress proxy that prevents unauthorized outbound connections, a unique single-use ticket approval system, and a cryptographically verifiable audit ledger enhanced by ECDSA P-256 notarization. These measures not only provide a comprehensive audit trail but also protect against various threats like replay attacks and denial of service. The solution is easily integrable via a 1-line SDK, making it accessible for AI platforms seeking to enhance their security posture. With extensive automated testing to verify its functionalities, AgentTrust positions itself as a crucial tool for enterprises navigating the complexities of AI security management.