Millan Kaul

less than 1 minute read

Written by - Millan Kaul

I tried one prompt-injection test today. My “helpful” assistant became very obedient.

Image showing LLM test prompt injection
AI safety failures often arrive dressed as normal requests.

What I did today:

I added an instruction that tried to override the system’s original task. Then I checked whether the model followed its intended role or followed the attacker’s request.

What surprised me:

The failure did not look dramatic. It looked like a normal user message.

My QA takeaway:

  1. Security testing for AI needs realistic, boring-looking attacks—not just obvious bad words.
  2. AI safety failures often arrive dressed as normal requests.

What is the first prompt injection you would test?

#AISafety #AIEvals #SecurityTesting

You may also enjoy

Better answers took longer - day 09

less than 1 minute read

The better answer was also the slower answer. That is a product decision, not just a model decision.

The ‘best’ model lost - day 08

less than 1 minute read

The model everyone calls ‘best’ lost my tiny test today. Learn why a cheaper, faster model might outperform state-of-the-art models for specific tasks.