← AgentsGitHub
AGENTIC ENGINEERING / SECURITY

Security follows the boundary.

An agent integration must constrain authority, not merely describe a task.

Security dimensions

DimensionBOUND question
AuthorityWho owns the decision and who may approve the result?
ScopeWhich files, resources, repositories, or records may change?
CredentialsWhich credentials are exposed, for what operation, and for how long?
NetworkWhich hosts or services are reachable from the execution context?
ExecutionWhich runtime, process, container, or sandbox contains the work?
ApprovalWhich actions require human or policy approval before execution?
VerificationWhich evidence must exist before changes are accepted?

Minimum bounded context

  • Explicit mission and owning boundary.
  • Allowlisted change surface.
  • Named tools and dependencies.
  • Credential scope appropriate to the task.
  • Resource and network limits where applicable.
  • Required verification commands or evidence.
  • Clear escalation path when the requested change exceeds the boundary.

Forbidden inference

An agent should not infer authorization from visibility, repository access, tool availability, or successful authentication. Those are capabilities or facts about the environment; they are not substitutes for an explicit boundary and contract.

Future implementation target: machine-readable permissions, sandbox policies, approval rules, and verification gates can later be attached to BOUND artifacts. This page intentionally does not claim that such a runtime exists today.