Page 383 - CyberSecurityEssentialsEnglish

⚠ A design rule
Treat a model’s output as untrusted input until it passes appropriate validation. Text that
looks natural may contain a command, a piece of code, a link, or a value that affects a
downstream system.

● Remember: Section summary
• Key idea: Risk sits in what the capability can reach and do, not in the name of the
product.
• Practical decision: Approve data types, limit permissions, set contractual requirements,

and treat every output as untrusted until it is validated.
• Mistake to avoid: Approving a tool once and leaving its permissions, its connections,
and its vendors unreviewed afterwards.

8.8 Law and Standards: An Obligation That Changes With Context and Time
◆ Key idea: obligations change with context.
There is no single global policy. An organization must tie its internal framework to local

law, sectoral requirements, the location of the people affected, and vendor terms, with
periodic review because the requirements change.

Voluntary frameworks and international standards help organize the work, but they do not

replace the law. An organization may be subject to rules on data protection, employment,
consumers, health, banking, or education, even where no standalone AI statute exists.
Obligations may also extend across borders when a service is provided to people in another

country.

The European Union’s AI Act entered into force in August 2024, and its provisions began to
apply in phases. Obligations on AI literacy and certain prohibited practices became

applicable from 2 February 2025, and rules on governance and general-purpose model
obligations began on 2 August 2025, while the remaining obligations were set to phase in
according to the type of system. That timetable has since changed.

A simplification package covering artificial intelligence received final approval in the Council
on 29 June 2026, following a European Parliament vote earlier that month. It defers the

obligations for standalone high-risk systems listed in Annex III to 2 December 2027, and for
artificial intelligence embedded in regulated products under Annex I to 2 August 2028. The

transparency obligations in Article 50, which cover telling a person that they are interacting
with a system and labelling synthetic media, were not deferred and applied from 2 August
2026. That last point matters most for the ordinary reader, because it is already in force. The

file continues to move, so official pages and current legal advice must be consulted [7].

Article 4 of that Act requires measures to ensure a sufficient level of AI literacy among staff
and others who use systems on the organization’s behalf, taking account of their experience,

382