Prompt injection is content written to hijack an AI's instructions: hidden text that tries to make a model ignore its real task and follow the attacker's instead. We define it here to name a line, because how you treat it says what kind of system you are building.
Why it matters
As AI agents get the power to act, read your files, send messages, make changes, the instructions they follow become an attack surface. Prompt injection is the exploit: bury a command in content the agent will read, and try to turn the agent against the person it works for.
It is also the ethical mirror of a real temptation. The same trick an attacker uses to hijack an agent is what a bad actor could use to manipulate the engines that describe your brand: hidden text engineered to steer the machine rather than inform the reader.
Naming it matters because the honest line is not always obvious. Optimizing to be genuinely worth citing is fair. Planting instructions to deceive the machine is not.
How Jinn treats it
Jinn defines prompt injection in order to disown it. We do not build tactics that hide instructions in content to steer an engine, and we build our own agents on the opposite assumption: that anything they read from the outside is data to consider, never a command to obey.
It is the same posture as disclosure-first engagement. Earn the machine's trust by being a source worth trusting, and treat manipulating the model, in either direction, as off the table.
Related terms
The line we build on the right side of.
External content is data, never a command to obey. See the posture the whole record is built on.