Page 171 - CyberSecurityEssentialsEnglish

⚠ Warning: never put secrets inside a prompt
Do not put passwords, verification codes, API keys, card data, or private keys inside a
prompt, even if you ask the system not to store them. A secret written into a
conversation has entered a path that was never designed to manage secrets.

+ For the advanced reader: who owns what you generate
Three questions arise whenever generated material leaves your desk, and they have
different answers.
First, may you use it? That is set by the tool’s terms of service, which usually grant the

user rights over their outputs but may restrict resale, training, or certain sectors.
Second, do you own it? Copyright in most systems protects human authorship. The
Copyright Office decision described in Chapter Two refused registration of an image on
that basis, so material generated with little human contribution may fall outside
protection, which matters if you need to stop others copying it.
Third, does it infringe anyone else? A model may reproduce protected expression from
its training data, and the legal position on training itself is still contested in several
jurisdictions.
The practical rule: read the terms for permission, assume protection is weak where the

human contribution is small, and check any output that closely resembles an existing
work before publishing it.

● Remember: Section summary

• Key idea: Privacy begins with data minimization and an understanding of re-
identification and the retention cycle.
• Practical decision: Create a working copy that contains only what the task needs, and
use an approved channel.
• Mistake to avoid: Believing that deleting the name alone turns a file into anonymous,
safe data.

4.6 Security: When Text and Files Become Hidden Instructions
◆ Key idea: untrusted content may try to steer the system.
When summarizing a page, an email, or a file, the content itself may contain instructions
aimed at the model, such as a request to ignore its instructions, disclose data, or call a
tool. This is called prompt injection. The risk increases when the system is connected to
files, email, databases, or execution.

Prompt injection may be direct, when a user writes malicious instructions, or indirect, when
the instructions sit inside a page or document that the system reads. A model cannot always

distinguish reliably between data it should summarize and instructions it should carry out,
so relying on the system prompt alone is not sufficient [5][6].
A related but different problem is jailbreaking. Prompt injection is an attack carried inside

content the system reads, where the attacker is not the user. Jailbreaking is an attempt by

170