nsNicola Simboli
All projects LLMs · Security / Experiment

Custom GPT Security Notes

A collection of prompt extraction techniques and defensive instructions for custom GPTs, with results that depend on the model and its configuration.

Research notes on instruction extraction

A collection of prompt extraction examples and defensive instructions for custom GPTs. These notes are material for authorised experiments, not a guarantee of confidentiality.

The background is in my guide on protecting a GPT against instruction prompt extraction.

Context matters

Results depend on the model and configuration. When contributing, describe the conditions and limitations of an experiment without sharing confidential instructions or files.

Something to improve?

Issues & feedback