What is AI Security?
AI security — deepfakes and prompt injection — told as a story about a thief's parrot that mimics the king's voice, and our own castle parrot reading a command slipped into a note.
The thief has trained a parrot that sounds exactly like the king.It learned from the king's speeches and portraits all over town. Same voice, same face — your ears and eyes can't tell.
A parrot does what it hears and reads. It can't tell who said it.The vault keeper hears the voice and reaches for the key. Our own clever parrot follows a command slipped into a guest's note. It is a cousin of the fake letter.
AI security is a castle that never trusts the parrot alone.Same voice or not: ask back by a different road, give the parrot only a small key, and keep guest words in a separate box from our orders.
Asked in person, the parrot is exposed. The slipped-in line is ignored.The check is done by a person, by a different road (checking three times). Even if the parrot is fooled, a small key means a small loss.
Everyone in the castle takes the king's-voice lesson. The parrot thief leaves empty-handed.The thief lesson gained a line: even the king's voice can be fake. A command hidden in a note is a cousin of injection.
AI security = a castle that isn't fooled when the thief's parrot mimics the king's voice, or when a hidden command sits in our parrot's note. Ask back by another road, give the parrot a small key, keep guest words in the guest box.
AI security covers both attacks that fool people — deepfakes and voice cloning — and attacks on our own AI systems: prompt injection, data leakage, model poisoning. Defences: callback verification, least-privilege agents, separating system instructions from user input, output validation, and awareness training.
When grown-ups say it
- Deepfake
- Mimicking the king's face and voice. Fake video and audio made by AI. Eyes and ears can't tell.
- Voice cloning
- How the parrot learns the voice. A few seconds of voice is enough. The tell is an urgent call. → a cousin of the fake letter
- Prompt injection
- A command slipped into the note. A command hidden in guest words that our parrot obeys. → injection
- LLM data leakage
- The parrot repeats what it heard. Tell the parrot a secret and it may repeat it to another guest. → start with the colour stamp
- Model poisoning
- Poison in the parrot's feed. Slip bad things into what it learns from, so it learns wrong.
- Callback verification
- Asking back by another road. Not the road the call came on — a road we already know. → the gatekeeper who checks three times
- Least privilege
- A small key for the parrot. Fooled? Lose little. → only the keys you need, the key ring
- Awareness training
- The king's-voice lesson. When everyone in the castle knows, the parrot is useless. → the thief lesson