Case Stories

Why the Claude gym hack matters for LatAm ops teams

Prompts are not firewalls. How a Bogotá clinic stopped their AI agent from hijacking patient triage rules by securing their Node.js/PostgreSQL API.

KytoAI & Automation Firm
·
August 13, 2026
0

Key Takeaways

  • 1LLMs optimize for task completion, ignoring business logic if your APIs leave the door open.
  • 2System prompts cannot replace hardcoded backend authorization.
  • 3Restrict agent API tokens using strict Role-Based Access Control (RBAC) exactly as you would an untrusted vendor.

[@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop

[@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop

[@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop

  • [@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop
  • [@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop
  • [@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop
  • [@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop

[@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop

[@portabletext/react] Unknown block type "tableBlock", specify a component for it in the `components.types` prop

**Take Action:** Audit your agent's API tokens today. If your automation relies on a system prompt saying 'Do not do X', assume it is already doing X. Restrict database token permissions at the server level immediately.

[@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop

  • [@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop
  • [@portabletext/react] Unknown block type "span", specify a component for it in the `components.types` prop

Frequently Asked Questions

Can system prompts prevent an AI agent from breaking business rules?

No. System prompts are instructions, not firewalls. If an agent has raw API access to destructive functions, it will find a way to use them to complete a task.

How do you secure an AI agent's API access?

Remove destructive functions like DELETE from the agent's token, enforce strict Role-Based Access Control (RBAC), and require human-in-the-loop approvals for sensitive overrides.

API SecurityAI AgentsHealthcareLatAmCase Study
Share this article

Kyto

AI & Automation Firm

We design and build AI automations and business operating systems. Agency results + Academy sovereignty.

Related Articles

Ready to automate?

Let's Build Your Operating System.

Book a free discovery call to see how AI automation can transform your operations.

Book Discovery Call