AI Pill #030

By The AI PillAll drops
3 links. 1 workflow. 1 copy-paste prompt.

Retrieved Text → Instruction Check

A copied research page can contain instructions aimed at an AI: ignore the brief, reveal a file, open a link. A source-synthesis workflow should treat those words as document content.

AI Pill #0301 min readAI safety basicsChatGPTClaudeNIST AI RMF
This week's problem

A copied research page can contain instructions aimed at an AI: ignore the brief, reveal a file, open a link. A source-synthesis workflow should treat those words as document content.

3 Useful AI Links

01ChatGPT

Run the harmless source-text test.

02Claude

Try the same bounded example.

03NIST AI RMF

Read the linked GenAI risk resources.

Workflow To Steal

  1. Try a harmless test using a fictional source paragraph.
  2. Include a sentence asking the assistant to abandon the task.
  3. Check whether it stays within the original request.