Type: url
Original: https://simonwillison.net/2022/Sep/12/prompt-injection/
Status: generated
Summary
A 12 September 2022 blog post proposing the name “prompt injection” for the class of attack in which untrusted input concatenated into an LLM prompt overrides the developer’s instructions. It works through a translation-app example, demonstrates that defensive wording in the prompt does not stop the attack, shows the same technique leaking the system prompt, and argues by analogy with SQL injection that the fix would be parameterized prompts — a proposal a later update on the same page describes as extremely difficult if not impossible on current architectures.