Architectural Prompt Injection Immunity at the Tool-Call Boundary

Why system prompts fail and how pre-action gates eliminate prompt injection at runtime.

The Failure Mode

Prompt injection cannot be solved by text instructions alone. Once an agent receives untrusted inputs and tool execution permissions, any prompt-layer defense can be bypassed with base64, nested translations, or recursive tool calls.

Architectural Resolution

ThumbGate segments the control plane from the data plane. Untrusted inputs are quarantined, and pre-action hooks validate every proposed tool call and argument against deterministic semantic schemas before execution.

// PreToolUse gate: enforce tool parameter boundary validation
function evaluateToolCallBoundary(toolName, args) {
  if (isDangerousSystemCommand(args) && !hasDeterministicApproval(args)) {
    return { decision: 'block', reason: 'PreToolUse: untrusted payload attempted unauthorized exec' };
  }
  return { decision: 'allow' };
}