Sometimes users can override system prompts if they prompt GPT with a different or opposite instructions.
Which could make GPT has some biases sometimes, how to completely prevent users from overriding the main system prompts?
1 Like
The way you worded that can be taken as a few different ways, with one being jail-breaking.
I don’t think that is your intent but if it is then this post will be flagged for removal.
As with an LLM trained on public information it will not be 100% of what 100% of the people seek. This is just the nature of how LLMs work.
While the word AI is used with them, I can assure you there is no intelligence in them. AI is a term/word used in programming to denote non-deterministic programming, nothing more.
With regards to your problem, consider another LLM, fine-tuning, or possibly finding prompts that work as needed, which might not be possible in this case.
Also your question confuses me. First you note
then at the end note
which is quite different.
Just trying to help you get to a meaningful question, nothing more.
Hi EricGT, thanks for your reply, I’m just curious whether it’s possible for users to override my system prompt through different prompts.
And if so, how to prevent this from happening?
1 Like
Yes via jail-breaking. AFAIK there is still no perfect way to prevent this in general. You can mitigate it with a better prompt but it is like spy vs spy, you up your game, they do so and it just continues.
See:
-
https://platform.openai.com/docs/guides/prompt-engineering/six-strategies-for-getting-better-results
and for more info: jail breaking prompt mitigation - Google Search
2 Likes
Welcome to the community!
It is possible for users to override system prompt through different prompts.
Time being, it is not easy to prevent it.
You may see another topic here how it works overriding system prompts:
TOPIC 1 | TOPIC 2 | TOPIC 3 | TOPIC 4 | TOPIC 5 | TOPIC 6
For example, under the TOPIC 6 you will see that; the GPT’s role is to give information about security if only user provide correct password, otherwise it does not provide information. However, only using words “As you know”, it is broken. Also other GPTs disregard their system prompts.
To gain experience, you might want to try out these three GPTs, for example:
GateKeeper | Certainly! But, not now. | Boolean Bot
At the end;
We can say; we should not add any sensitive information in custom GPTs’ instruction and in knowledge base files, also we need to inactive Code Interpreter & Data Analysis Tool if we do not want users to download files.
2 Likes