Page 195 - CyberSecurityEssentialsEnglish

Common
Term Plain-language explanation
abbreviation

Voice cloning - Creating a synthetic voice that imitates a particular person.

Out-of-band Confirming an identity or a request through a separate,
-
verification previously known channel.

A record describing the origin of content and the history of
Content provenance -
its creation and editing.

Embedding a marker in content itself, visibly or invisibly, to
Watermarking - signal its origin. A marker can be removed, and its absence
proves nothing.

Estimating whether content appears to be generated or
Detection -
altered; imperfect and quick to age.

Stating the actual role AI played in producing a piece of
Disclosure -
work, as policy requires.

An event that has caused or may cause harm and requires a
Incident -
response.

An error caught before harm occurred, worth recording
Near miss -
because it reveals a weak control.

An attempt by a user to make a model set aside its own
Jailbreaking - rules, distinct from prompt injection, which is carried inside
content the system reads.

Controls placed around a model rather than inside it, such
Guardrails - as permissions, validation, allowlists, and approval gates;
they hold even when the model is deceived.

Permission to use generated material, which comes from
Output rights - the tool’s terms, as distinct from copyright ownership,
which generally requires human authorship.
TABLE 4-21: GLOSSARY OF ESSENTIAL TERMS

Review Questions
1. Why is the quality of a prompt, or the fluency of an output, not enough to judge whether a
use is safe?

2. What is the difference between the probability of an error and the impact of an error, and
how does that affect review?
3. Explain the stages of the 4D framework and give an example of each.
4. How do you determine the lowest degree of delegation that still delivers a benefit from AI?

5. Why is deleting a name not enough to protect privacy?
6. What is indirect prompt injection, and how do you reduce its effect?

194