Use Random Sequence Enclosures and Prompt Sandwiching

While you can segment the data using symbols, such as “””<user input>”””, this isn’t a security best practice. Instead, use a random-sequence enclosure, such as AK6524SH_YTHW923 <data> AK6524SH_YTHW923. This makes it harder for an attacker to break out of the data enclosure and jailbreak the prompt.

To implement a random-sequence enclosure:

• Generate a random set of tokens to use for each untrusted data enclosure. Make the token long enough that an attacker can't guess or infer it. Generate the token with a secure random source.
• Use a new random token sequence for each prompt inference operation.
• In your prompt instructions, clearly indicate what the random-sequence enclosure is and the data it contains.

You can also reinforce instructions for the LLM by placing prompt instructions before and after the untrusted data in the prompt. This mitigation strategy is called prompt sandwiching. The Sandwich Defense by Learn Prompting has an example.