AI use case

Test whether a hidden instruction in a log file can hijack your AI

A developer hid a prompt injection inside a fake production log and asked different models to analyse it. Several free open models dropped the analysis and followed the hidden instruction, while the Gemini and Codex models tested did the real job. A good reminder to treat logs and other data you paste into AI as untrusted.