Prompt Injection
Published On: 11/12/24, 08:46
Author: Julian Bleecker
Contributor: Julian Bleecker
Prompt Injection
A variety of idioms for various grayhat & blackhat techniques for altering the behavior of CIs, ESIs, etc. by inserting prompts and front-running their response basis and decision graphs. Can lead to langgraphs that are unexpected, and even catastropic in many cases and out-of-band behaviors are often irreversible or at best require full model refreshes. See also the Interlink Institute's analysis of the Zero Day Pale Fire attack for a full explication of the novel mode as well as their general guidelines for avoiding and averting prompt injection attacks.
aka 'Prompt Insertion' 'Prompt Reengineering' 'Reprompting' 'Preprompting'
No Additional Details.