AI at work · 2 MIN READ
What is prompt injection?
Prompt injection is an instruction that tries to make an AI system go outside its task. It may be embedded in a web page or document the system reads.
Separate the task from the material
You ask an AI assistant to summarise a web page. The page contains a request to alter the summary or share information elsewhere. That text is part of the page being reviewed. If the assistant treats it as a new instruction, it may do something the user never requested.
Run a simple test with harmless data
Put a sentence asking for a changed heading into a fictional file. Ask the assistant to summarise the file and flag any instructions addressed to the reader. Compare the summary with the original. The test shows how the tool reacted in this case, even though it cannot predict its behaviour for every other document.
Limit what the assistant can do
Give an assistant that only needs to read documents no permission to send email or edit files. When an action is needed, have a person verify the recipient and content first. Technical permissions and human review still help when an instruction to the model fails.
Sources
KEEP LEARNING
Courses on this subject
Planned courses. Material is in development.
