
Wise AI Ambassador · Set 17 of 20
Protection
Cards for the fourth degree of the Wise AI Ambassador
Tell technical attacks on systems from cognitive attacks on people, and build a shield against each.
A card for each unit of the degree, with its idea, its exercise and its takeaway, and the gate's card at the end.
- Degree 4 · Protection1From information security to AI security
A model is a new attack surface.
ExerciseMap where data enters an AI system you use, and where it leaves.
TakeawayEvery input to a model is a door that needs a guard.
Fawzoozfawzooz.ai - Degree 4 · Protection2Five attacks you must know
Know the attack to know its shield.
ExerciseChoose one attack, and write a realistic scenario for it in your organisation.
TakeawayWhat you cannot name, you do not see when it happens.
Fawzoozfawzooz.ai - Degree 4 · Protection3Red-teaming models
Attack your system before someone else does.
ExerciseWith permission and in a test environment, try to make an assistant break its instructions, and record what happened.
TakeawayA flaw you find yourself costs less than one someone else finds.
Fawzoozfawzooz.ai - Degree 4 · Protection4Cognitive security
The human mind is the easiest target.
ExerciseLook back at the last urgent message you received, and pick out the pressure it put on your feelings.
TakeawaySlow down: haste is the attacker's first weapon.
Fawzoozfawzooz.ai - Degree 4 · Protection5Deepfakes and identity
A voice and a face are no longer proof.
ExerciseAgree a verification word, or a second channel for confirming, with your team or your family.
TakeawayCall back yourself; don't rely on the one who called you.
Fawzoozfawzooz.ai - Degree 4 · Protection6Quantum and post-quantum cryptography
Prepare for encryption that may break tomorrow.
ExerciseAsk your technical team: which of our data must stay secret for ten years?
TakeawayWhat must stay secret for a long time must be protected from today.
Fawzoozfawzooz.ai - Degree 4 · ProtectionGate 4 · The output
An adversarial testing report
What you doTest at least one technical attack and one cognitive attack, in a safe, authorised environment.
You have passed itDo not move on to the next degree before you pass this gate.When you record what you found and its impact, and propose a shield for each.
Fawzoozfawzooz.ai
