The Agent Access Model

The Agent Access Model (AAM) proposes extending Zero Trust principles to AI agents, whose short-lived, machine-speed tasks and multi-system access needs make traditional human- and service-account controls inadequate. Rather than trusting a task after initial authorization, AAM evaluates every action against the agent’s identity, the authorized task, and the task’s accumulated security state, with capabilities only able to narrow over time. Its reference architecture combines short-lived, task-scoped, sender-constrained credentials; a task-scoped access engine; enforcement in both the agent harness and network egress layer; a Trust Ratchet that removes permissions after protected events such as sensitive-data reads; and logging and review systems for refining future permissions. The model emphasizes that prompts cannot enforce security boundaries, since agents may be manipulated by injected content, and that enforcement must occur where tool calls and network requests are actually executed. A finance reconciliation example illustrates how a protected report can trigger restrictions that block later attempts to send data to vendor support while allowing a narrowly defined internal summary. Activity records collected from external enforcement points support auditing, incident investigation, and evidence-based changes to task templates without expanding permissions during an active run. Human approvals are intended for exceptional, high-risk decisions rather than routine actions, reducing approval fatigue. The paper identifies shared or multi-user agents as an unresolved problem because data, context, cached outputs, and generated responses may carry differing permissions across users, requiring stronger end-to-end provenance and authorization mechanisms.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.