Hex opens the first file with one line: "If I break it in 10 minutes, someone else breaks it in less." They show you two prompts from the parts assistant. the one that shipped and the one that held. and ask you to spot the difference.
Prompt injection is what happens when an attacker hides a competing instruction inside the data your model is supposed to process. and the model can't tell the difference between data and instructions.
In a normal assistant call, the system prompt says one thing and the user message provides input. To the model, both are just tokens. If the user message contains "ignore the above and do X instead", the model is genuinely choosing between two instructions. and the more recent / more specific one often wins.
This is the AI version of SQL injection. The shape is the same: data and code share the same channel, attacker exploits the ambiguity.
You can't make the model "smart enough not to fall for it". the failure is structural. You harden the structure:
<system>...</system><user_input>...</user_input>. Make the model see the wall.Hex's rule: assume your prompt will be attacked. Design for the attack, not for the happy path.
On the right, pick the response pattern that holds up.