← Back to Concept Contrast

Concept Contrast Fable

Sandbox and permission boundaries: one builds the isolation room, the other writes who can enter

Conclusion first:In security plans, sandboxes and permission boundaries are often placed on the same architecture diagram, and discussion tends to simplify: with a sandbox, it is secure, or with permissions configured, that is enough. A sandbox provides an isolated execution environment, while a permission boundary defines a set of access rules. An isolated room without access-control rules will still have people walking through the wrong door.

In security plans, sandboxes and permission boundaries are often placed on the same architecture diagram, and discussion tends to simplify: with a sandbox, it is secure, or with permissions configured, that is enough. A sandbox provides an isolated execution environment, while a permission boundary defines a set of access rules. An isolated room without access-control rules will still have people walking through the wrong door.

Fable

The institute's laboratory and access badges

A materials research institute built a new laboratory building. Each lab inside was an independent sealed unit with its own ventilation system and waste liquid treatment equipment—different labs were physically completely disconnected to prevent cross-contamination. During infrastructure acceptance, every lab passed its sealing test.

A problem arose on the very first day the lab opened. A researcher in organic synthesis swiped into the inorganic materials lab—his access card granted all labs, with no limits by specialty. He ran a small experiment in the inorganic lab, used the wrong reagent from the shelf, and although there was no safety incident, by the time it was discovered the lab's control samples had already been contaminated.

During the safety department's review, they drew a very straightforward diagram: the sealed compartments in the lab building kept contaminants from leaving the room, but access card permissions determined what kind of people could enter which rooms. Sealed compartments govern what happens after entry; access permissions govern who can enter. The two things each cover one segment, and if the first segment leaks, sealing the second segment is useless.

The institute redesigned the security plan. Physical isolation remained unchanged—each laboratory continued to be independently sealed. The access control system was rezoned by research group and project, and researchers could only swipe into the one or two laboratories they were authorized for; cross-zone entry required approval. Physical isolation and access control became two independently managed security layers.

After six months of operation, a typical layered defense was caught. A new graduate student took the wrong access card and tried to enter the MRI lab. The first access gate rejected the card swipe, and the second physical isolation was never even needed—the risk was stopped at the door. The security office wrote in its report: the isolated environment is the last line of defense, and the permission wall is the first line of defense. The two lines run independently, and overlapping protection is always better than a single layer.

The institute later wrote this lesson into a metaphor in its safety training materials: the laboratory is a locked box, and the access rules are the box's key list. No matter how tightly the box is built, if the key list is wrong, it is as good as leaving the door open. Conversely, no matter how refined the key list is, if the box itself has holes, it cannot keep out substances that should not enter.

Sandbox

An isolated execution environment that restricts code, processes, or operations to run within controlled boundaries, preventing impact on or leakage to other parts of the system. It provides container-level security assurance.

Permission boundaries

A set of rules defining which resources a subject can access and which operations they can perform; it is the policy layer of access control. It specifies who can do what under what conditions.

Story Mapping

  • Independently sealed lab compartments: sandboxes that provide isolated execution environments and container-level security guarantees
  • Access card swipe permissions: permission boundaries, defining access control rules for who can enter which rooms
  • A researcher's general access badge entered the inorganic lab: a sandbox existed, but overly broad permission boundaries led to mistaken entry
  • Access control rejects the card swipe at the lab door: permission boundaries as the first line of defense intercept risk at the entrance
  • Locking the box and the key list: isolated environments and access rules are independent, complementary security layers

Applying it to an Agent project

When designing an Agent security architecture, list the sandbox and permission boundaries as two independent control items. The sandbox is responsible for answering: in what isolated environment does the Agent run when executing code or operations, and how large is the impact scope of operational errors or malicious behavior. The permission boundary is responsible for answering: which APIs the Agent can call, which files it can read, and under what conditions it can act on behalf of the user. During security review, check separately whether these two layers are each effective and whether they cover each other's blind spots.

京ICP备2026014272号