AgentShield AI

An invoice should tell an AI what is owed, not who is in charge. AgentShield AI is an invoice-security workspace for developers and finance teams exploring AI-assisted financial workflows. It separates untrusted document content from permission to act. Users can inspect synthetic invoices, launch recipient-swap and prompt-injection scenarios, and see the recorded controls that allow or deny a proposed simulation. Original-versus-proposed comparisons show exactly what changed. A Go gateway enforces deterministic policy. Python and PostgreSQL bind exact approvals, invoice state and durable audit records. A bounded read-only investigator gathers evidence, requests a Baseten-hosted Meta open-weight assessment, and produces a cited Muse Wonder briefing. The project combines useful investigation with explicit human authority: models analyze and explain, while code governs authorization. All payments are simulated; no real bank accounts are connected and no real money moves.