[DISCUSSION] Can AI actually be hacked?

Things I've been researching:

Prompt injection — can untrusted input manipulate an AI's instructions?

Jailbreaks — how do people test the limits of a model's safety behavior?

RAG poisoning — what happens when an AI retrieves misleading or malicious documents?

AI agents — what if a model has access to tools, files, or external services?

Data leakage — can a system reveal information it was never supposed to expose?

Model extraction — how much can someone infer about a model through repeated queries?

I'm looking for people who understand the difference between breaking a demo and discovering a real vulnerability.

Question: What's the most overlooked weakness in modern AI systems?

Not looking for recycled jailbreak prompts. Interested in actual security
research.