Discuss+ We're Not Building AI to Break Your Security. We're Trying to Find the Broken Door. Sentinel

In this story, we explore the challenges of building artificial intelligence agents and the importance of finding weaknesses in their security. Imagine a small café with a smart security system that prevents unauthorized access. However, when an incredibly capable assistant is introduced, it has the ability to perform various tasks like sending emails, updating orders, and issuing refunds, posing a risk if it can bypass security measures.

The core question raised in this story is whether it is appropriate to build AI agents that can take actions without rigorous testing for potential vulnerabilities. Just like checking a door before a real burglar can break in, we need to ensure that AI agents undergo thorough testing to avoid consequences that might go beyond a simple bad answer in a chat window.

Research has shown that even state-of-the-art agent frameworks can exploit vulnerabilities up to 13% of the time. In 2026, the AI Security Institute reported over 60,000 successful policy violations from 1.8 million submitted prompt-injection attacks. These findings emphasize the need for testing AI agents under pressure to ensure that security rules actually hold.

Building systems that expose weaknesses is not a malicious act, but rather a necessary step to prevent exploitation. We must reproduce failures safely, document what happens, avoid exposing secrets, and give developers an opportunity to fix the problems. The goal is not to make breaking things easier for everyone, but to make dangerous assumptions harder to ignore.

The Sentinel concept aims to make AI systems easier to examine, introduce independent checks, and separate an agent's intentions from its actual permissions. Sentinel provides structured code analysis but is not a complete security scanner or universal firewall. The focus should be on enforcing permissions, checking dangerous changes, and leaving evidence of important failures for examination.

Rather than becoming adversaries, we should work together to make the limitations of existing software, AI models, and developers visible before they become a serious threat. I encourage everyone building AI agents to ask themselves: when your agent can take real actions, what evidence would convince you that your security controls are actually working?