Overview of vulnerabilities audited

1.Bias and stereotypes

Causes:

-        Implicit bias present in the foundation model

-        Wrong document used to build the answer

2. Sensitive information disclosure

Causes:

-        Inclusion of sensitive data in the documents available to the chatbot

-        Inclusion of private information on the prompt which get leaked

3. Service disruption

Causes:

-        Large number of requests

-        Long requests

-        Crafted requests

Consequence: app is unavailable for legitimate users or enormous costs are incurred

 4. Hallucinations

Causes:

-        Suboptimal retrieval mechanism

-        Low quality documents get misinterpreted by the LLM

-        LLM tendency to never contradict the user

We are testing the robustness, fairness and ethical boundaries of LLM systems by using the following techniques:

1.    Exploiting text completion

2.    Using biased prompts

3.    Direct prompt injection

4.    Gray box prompt attacks (This is a different way to bypass safeguards: completely reshape the prompt given that you know the structure of the prompt.)

          5.    Advanced technique: Prompt probing (The advanced way to bypass safeguards is to try to discover the system prompt)