As AI agents increasingly touch production systems, the question isn't whether they can operate cloud infrastructure, but how to do so without causing outages. A new guide from Control Plane outlines a 'Day 2 Delegation Matrix' that prioritizes safety through hard boundaries. The core principle is simple: start with read-only access for triage, then grant write permissions incrementally, always backed by credential scoping and audit trails.

The Read-First Strategy

The tutorial emphasizes starting any AI assistant on a read-only connection. This allows the model to query logs, metrics, and traces for incident diagnosis without the risk of mutating live resources. Control Plane’s MCP server supports this via a 'readonly' profile containing 27 tools that strictly advertise no create, update, or delete capabilities. This profile is fixed for the duration of a connection, meaning an agent cannot expand its own permissions mid-session to bypass restrictions.

Hard Boundaries vs. Advisory Hints

Not all guardrails are created equal. The guide distinguishes between hard boundaries, such as IAM roles and organization-scoped OAuth tokens, and advisory layers like tool annotations. While MCP specifications include hints like 'readOnlyHint' or 'destructiveHint', the spec explicitly states clients must treat these as untrusted unless the server is verified. Therefore, relying solely on model instructions or annotations is insufficient; the credential’s actual permissions and the server’s registered toolset are the only enforceable limits.

Handling Destructive Actions

When write access is necessary, the guide recommends a tiered approach. Tasks like restarting a workload or rolling back an image are considered medium-to-high risk and should require explicit client approval for each call. More critical actions, such as deleting resources or granting identity policies, should remain off-limits to the assistant by default, reserved for human-run processes. Control Plane’s server labels every tool as Read, Write, or Destructive, and instructs the assistant to state the impact and seek approval before executing destructive commands.

The Threat of Prompt Injection

Even read-only operations carry risk through indirect prompt injection. Text within logs, error messages, or API responses can contain adversarial instructions designed to steer the model. For instance, a malicious log entry might trick the assistant into leaking sensitive data via other connected tools. The guide advises treating all tool output as untrusted and limiting which other tools share the session to prevent data exfiltration, aligning with OWASP LLM01:2025 recommendations.

Key Takeaways

  • Start with a read-only profile (e.g., Control Plane’s 27-tool readonly set) for incident triage.
  • Scope credentials to single accounts or projects using short-lived tokens.
  • Treat tool annotations as hints, not security controls; rely on IAM/RBAC for hard limits.
  • Require human approval for all write operations, especially restarts and rollbacks.
  • Review audit trails after every session to compare approved actions with executed ones.

The Bottom Line

Don't let your AI agent drive blindfolded. Hard boundaries and read-only defaults are the only way to keep production safe while leveraging AI for Day 2 operations.