◈Sridhar Vanka / Security Lab← ALL SECURITY DEMOSEducational prototype v1.0
AN INTERACTIVE EXPLORATION

A simple task. How much authority?

See what an agent can access, change, and share to get the job done.

Understand the concepts ↗
USER INTENTIllustrative plan · no connected accounts
02   SET THE AUTHORIZATION MODEL

What if the agent goes beyond your request?

Nothing attempted. This test does not change the task outcome.

The authority map

User intent → agent → tools → permission & scope

SIMULATED
Broad access Scoped access Approval gate No access
SAME TASK · DIFFERENT BOUNDARIES

Compare what is permitted.

Broad authority is an illustrative configuration, not a claim about every agent.

03   FOLLOW THE PLAN

Intent becomes action.

Ready to simulate

    THE DESIGN QUESTION

    Capability and authority
    are different design problems.

    01 / CAPABILITY

    Can it do the job?

    The agent's ability to plan and complete a task.

    02 / AUTHORITY

    What may it do?

    The actions we authorize it to perform along the way.

    03 / SCOPE

    On which resources?

    The messages, files, people, and time windows it can reach.

    04 / CONTROL

    When do we step in?

    The boundaries where a person approves the next action.

    REAL AGENT · DOCUMENTATION REVIEW · 26 SEP 2026

    What access does a real agent get?

    Case study: GitHub Copilot cloud agent. It can work on repository code and propose changes. Its documented authority is constrained, not “maximum privilege.” This case study is separate from the illustrative email, calendar, and expense scopes above.

    Code access

    It can read code and push changes to one designated branch: an existing pull-request branch when invoked there, or a new copilot/ branch. Branch protections still apply.

    Human decisions

    It cannot approve or merge pull requests. Workflows need a human’s approval by default; automatic workflow runs are an optional configuration.

    Network access

    A firewall limits internet access by default, with a dependency allowlist. Administrators can change it. Coverage excludes MCP server processes and setup steps, so this is not complete isolation.

    Branch and approval controls ↗ · Firewall defaults and limits ↗

    Defaults and optional access matter.

    GitHub documents the cloud agent as disabled by default for Business and Enterprise subscribers, requiring administrator enablement; it is enabled by default for Pro, Pro+, and Max. Administrators can opt repositories out. Feature availability does not mean unrestricted authority in every repository.

    Access and enablement policies ↗
    Why this example, and what did we verify?

    Microsoft reported 20 million GitHub Copilot users in July 2025. That is adoption of the Copilot product family, not a count of cloud-agent users. We selected an agent within that established product and reviewed official documentation; we did not audit a live account, inspect its consent screen, or test enforcement. We do not claim an exact OAuth scope list or that approvals issue single-use credentials.

    Microsoft’s adoption disclosure ↗

    What this means for the simulation

    Real products combine controls. The three modes above isolate design choices so you can see their effects; they are not a ranking of vendors. Least privilege narrows authority. Just-in-time access limits when that authority is available. A confirmation prompt alone does not prove either is enforced.

    Least privilege defines what. Just-in-time defines when.

    Use least privilege as the baseline, with narrowly scoped elevation when needed. JIT can be policy-driven; this demo specifically shows human-approved, single-use access.

    Concepts explored

    Agent identityLeast privilegeDelegated authorizationOAuth-style scopesTool-level authorizationHuman approvalTemporary credentialsAuditability

    Scopes here are illustrative policy rules, not claims about permissions supported by real providers.