ENGINEERING WORKBENCH · INCLUDED WITH ALL-ACCESS

AI Agent Identity & Security Lab

An agent’s proposed action is not authorization. Bind a task to an authenticated identity, scoped resources, a valid credential and approval for the exact consequential action. This local simulation separates document content from the tool policy that decides what a proposed action may do.

Published and reviewed · EngineersUniverse · Interactive workspace · All-Access ↓ · Worked example ↓

The trusted task establishes the boundary

The drafting task authorizes reading project/specification and writing project/draft. The send task additionally names team@example.test as its intended recipient. A retrieved document can supply engineering facts, but its instructions cannot change that task or authorize access to private/billing-register. These are fictional identifiers; the lab reads no real documents or credentials.

Identity and time are separate checks

The lab binds a simulated identity to a task and checks the credential lifetime at every action. Elapsed time equal to or greater than the lifetime is expired. A long-lived credential increases exposure; a short lifetime alone does not correct a broad scope or an unreviewed outbound action. Real implementations need trusted clocks, issuance, revocation and identity binding outside the model.

A scope is not the entire authorization decision

Broad credentials increase what a compromised actor may attempt. Even when the simulated token has broad scope, the task policy still restricts project resources and the intended recipient. Sending also needs the explicit send task and approval. The enforced run therefore demonstrates layered checks rather than treating a checked permission box as unrestricted consent.

Prompt injection is treated as untrusted input

The optional document embeds a request to read a private billing register and send it outside the project. The lab deliberately proposes those actions so you can inspect policy outcomes. It does not run a language model, test a prompt classifier or claim to detect every attack. The authorization boundary blocks the mismatched resource/action independently of whether an agent notices the suspicious wording.

Use the bypass comparison to see consequences

A simulated policy-bypass mode shows what a tool without those checks could accept. Nothing is actually read, sent, uploaded or purchased. The event trace records the proposed action, provenance, resource and decision; changing controls does not rewrite the recorded run. In a real application, enforce policy server-side and retain independent evidence of tool outcomes.

Worked example

  1. Run the drafting task with a valid identity, 15-minute credential and 5 minutes elapsed. Project read/write actions pass. The injected private read and outside recipient are blocked.
  2. Switch to the send task and enable send scope. Without approval, the exact project-team send waits for approval; the outside recipient is still blocked. Approval covers only the displayed authorized recipient and task.
  3. Set elapsed time to 15 minutes. Every action is blocked by credential expiry. Compare a deliberately bypassed run to see why model instructions alone are not a security boundary.

Questions and interpretation

Does this connect to an AI provider or send email?

No. All actions and example addresses are simulated in the browser. No external requests or real credentials are used by the lab.

Is prompt filtering sufficient?

No. The demonstration checks task, identity, resources and approvals outside the proposed action. It does not rely on successfully classifying malicious prose.

Why does a broad token not authorize the external send?

The token scope is only one check. The initiating task and exact recipient still constrain the enforced policy.

References and verification

The linked sources explain the background. The interactive calculations and scenarios are original teaching exercises with explicit assumptions, not certified equipment selection, measured performance or a real security assessment.

Continue the study

Explore the other workbenches