Unreliable input filter in AI penetration test
🤯 Sometimes an AI chatbot refuses to answer a question. The question gets blocked. When I get it to answer exactly the same question, it can get dangerous! I recently discovered exactly this during an AI security penetration test. 🔍 I'll take you behind the scenes of an ethical hacker's work. One of the things I tested was the chatbot's input filter. Think of it as a security checkpoint: its job is to inspect questions before they reach the AI model and block requests that violate predefined security rules. To test whether this protection worked reliably, I submitted a question that should trigger the filter. Here's what happened: - First attempt: 🚫 The input filter recognized the question and blocked it. The chatbot refused to process it. - Second attempt: ✅ I submitted exactly the same question again. This time, the filter didn't trigger, and the question was answered. Same question. Same chatbot. Different security decision. 💡 Why is this a problem? An inpu...