Common advanced LLM powered systems can have severe security risks.
Common security risks are:
-
AI systems can malfunction when exposed to untrustworthy data, and attackers are exploiting this issue.
-
New guidance documents the types of these attacks, along with mitigation approaches.
-
Prompt injection.
-
Leakage of personally identifiable information (PII)
-
Harmful prompts. Relevant when you develop your ‘own’ LLM or LLM powered application.
No foolproof method exists for protecting ML/AI systems from security hacks. This is problematic when ML/AI systems are used for health systems, transport systems or weapons. Misdirection is a common threat.
Using LLMs for health saving systems or software that is used for safety applications (cars, trains, plains):
:::{danger} The outcome of LLMs should never be trusted. Despite imense progress on LLMs and their applications: Transparency is often absent and outcomes should never ever be trusted without a SOLID and GOOD human assessment! :::