DEVOURED
The Agent Access Model
Cloudflare proposes the Agent Access Model (AAM) to replace implicit trust in AI agents with task-specific, ephemeral credentials and strict harness enforcement.
What: The Agent Access Model requires that every action taken by an agent be authorized based on the current task's state. It uses five core principles: short-lived credentials, enforcement in the harness (not the prompt), minimal human oversight, evidence-based grant reviews, and a 'Trust Ratchet' to reduce capabilities as a task progresses.
Why it matters: This addresses a critical security gap where current access control systems (built for humans) fail to protect against AI agents that act at machine speeds and consume broad permissions that persist after tasks conclude.
Takeaway: Start by instrumenting a single agent task to use short-lived, task-scoped credentials and route its tool calls through a central harness that enforces granular allow-lists.
Deep dive
- Agent Identity Broker: Mints task-specific, sender-constrained tokens that expire when the work is finished.
- Task-Scoped Access Engine: Validates every request against a predefined 'capability ceiling' assigned at task dispatch.
- Mediation Layer: Intercepts tool calls and network egress to prevent unauthorized actions.
- Trust Ratchet: A stateful mechanism that programmatically reduces an agent's authority as it moves through a task flow.
- Grant Review Loop: Uses logs of actual activity to refine access permissions, preventing over-privileging while maintaining security.
- Multiplayer problem: Acknowledges that agents acting on behalf of multiple users with different permissions remains a difficult, unsolved challenge.
Decoder
- BeyondCorp: A security framework that assumes the network is hostile and authorizes access based on identity and device posture instead of network location.
- Sender-constrained token: A security token that is cryptographically bound to the client, preventing it from being intercepted and replayed by attackers.
- Harness: The runtime environment that brokers tool calls and network traffic for an AI agent.
- SIEM (Security Information and Event Management): Software that collects and analyzes security logs from across an entire organization.
Original article
Full article content is not available for inline reading.
Explore all 58 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi