Skip to lesson
Exit
Model, MCP, Integration and Safety Controls1 / 3

1 min lesson

Admin control plane

Use "You can sort each governance control into steer-only or enforcing and name the mechanism that enforces a risky operation" to say what you would do next.

Step 1 of 3

The case: you can sort each governance control into steer-only or enforcing and name the mechanism that enforces a risky operation. Start with this check: Model and MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. access match policy, and Auto's routing preferences leave only the optimization modes policy allows. Walk it once against a real situation so you can defend the call, not just assert it.

Learn more

Full explanation

Admin control plane

Enterprise controls — stop agents running destructive terminal commands
Admin control plane
SayPlatform security wants one thing: agents must never run a destructive shell command like rm -rf on a developer's machine. The tempting fix is a rule that says 'don't.' I sort the available controls into what steers and what actually enforces before I pick.
SayA Team Rule or an AGENTS.md line like this steers the model. It is guidance, and the model can still ignore it or work around it.
Typenever run rm -rf or other destructive shell commands
SeeUseful as a baseline, but non-deterministic. On its own it does not stop the command.
DoAdd the deterministic layer. Set the org Run Mode policy to Auto-review instead of Run Everything so shell commands are sandboxed and risky calls need approval, and distribute a hook at the agent lifecycle point that blocks the command outright.
SeeNow the boundary is enforced by the sandbox and the hook, not requested by a rule. The rule still steers; the Run Mode policy and the hook enforce.
DoAudit that the control changed behavior, not just that it was configured. Try a destructive command in a test repo and read the audit log.
SeeThe command is blocked and logged. If it had still run, the rule was carrying weight it can't hold and the enforcement layer was missing.
SeeThe proof point: the risky operation is stopped by a deterministic control, the rule is named as steering only, and the audit shows behavior changed. Model instructions were never treated as hard policy.
Learn more

Optional practice

Practice: Admin control plane

QYou've applied it. What actually proves the work is done, not just a plausible answer?