- Definition: users intentionally trick the LLM into doing something it shouldn't, which can lead to biased, offensive, dangerous, or harmful outputs.
1:48 - Real cases
1:48:- A Chevrolet dealership chatbot "sold" a car for $1 after a user manipulated it into agreeing with every request.
- A DPD chatbot swore at a customer and wrote poems mocking the company after being prompted to — which went viral on social media.
- Distinction from hallucinations: jailbreaks are adversarial (deliberate manipulation), not accidental model error.
Mitigations covered later: guardrails and adversarial red-teaming.