A document-management system let its internal AI engine turn free text into internal API calls. The result was IDOR – except the input came from a model, not from a user.
The Chatbot That Called Our Admin API: LLM Function Calls Without Access Control
A support chatbot was wired straight into the backend with no access-control layer between them. Asking it politely for a little extra produced a call to /api/admin/exportAllUsers.
Prompt Injection to Account Takeover: A Real Chain Through an AI Feature
A breach that chained three small weaknesses: blind use of an LLM, SSRF through prompt injection, and a weak JWT. Individually minor; together, full system access.
“But We Asked the AI for Secure Code”: Five Minutes, Three Vulnerabilities
A development manager told me proudly they now build five times faster with AI. I asked about security. Five minutes later I had found three vulnerabilities in the AI’s code.
How Prompt Injection Actually Works, in Two Requests
User input gets planted inside a prompt the server sends to the model. From there the road to prompt injection is very short – and MathGPT’s API key is what it looks like when someone walks down it.