Kavida.ai — agentic threat model

Not certain from the listing — Agent PO likely relies on commercial foundation models (e.g., GPT-4, Claude) to parse procurement requests and generate purchase orders. Threats include prompt injection leading to unauthorized purchases or model misalignment causing incorrect order quantities.

Not certain from the listing — The agent must access procurement databases, supplier catalogs, and ERP systems. Threats include data poisoning of supplier catalogs or exfiltration of sensitive pricing and vendor data via prompt injection.

Not certain from the listing — The agent likely uses a custom framework to plan procurement workflows, generate purchase orders, and call ERP APIs. Threats include insecure tool integration (e.g., executing arbitrary API calls) and memory poisoning from malicious supplier emails.

Not certain from the listing — As a closed-source SaaS procurement tool, it likely runs in a cloud environment. Threats include insecure storage of ERP API keys, lack of network isolation, and potential privilege escalation if the agent's container is compromised.

Not certain from the listing — Procurement agents require strict audit logs and human-in-the-loop (HITL) approvals. Threats include blind spots in logging agent-initiated transactions and a lack of drift detection in automated purchasing patterns.

Not certain from the listing — Procurement systems require strict role-based access control (RBAC) and financial compliance (e.g., SOX). Threats include unauthorized purchase authorization if the agent lacks robust identity verification and session management.

Not certain from the listing — While primarily interacting with ERPs and suppliers, future scaling could involve multi-agent negotiation. Threats include rogue supplier agents manipulating Agent PO into unfavorable terms or cascading transaction failures.