Page 157 - CyberSecurityEssentialsEnglish

“Seeing faces and hearing voices is no longer enough. A high-impact request needs an
independent verification channel and a decision record.”

◆ Key idea: safety is the management of control.
Using AI safely does not mean fearing every tool or reviewing every word in the same way.
It means matching the level of protection and verification to the sensitivity of the data, the
impact of error, and the permissions of the system, so that a person remains able to

Chapter Objectives
Apply the 4D framework: define, delegate, discern, and document.
Determine the impact of error and the level of risk before entering data or accepting
outputs.

Know what should not be uploaded to public tools, and how to minimize data and remove
identifiers.
Understand security risks such as prompt injection, malicious documents, improper output

handling, excessive agency, and the risks of agents connected to tools.
Verify claims, figures, and sources using primary sources, lateral reading, and the SIFT
method.

Distinguish harmful bias from natural variation, and test fairness and representation in
practical use.

Handle deepfakes, voice cloning, and synthetic content while distinguishing detection,
provenance, and verification of the claim.
Document the use of AI, the human changes made, and the decisions taken, in a light and
practical way.

Carry out a correct first response when a leak, a wrong decision, forged content, or
unexpected behavior is discovered.

understand, stop, correct, and be held accountable.

FIGURE 4-1: THE 4D FRAMEWORK FOR SAFE AND PRODUCTIVE USE

156