Page 379 - CyberSecurityEssentialsEnglish

The Contract as a Governance Instrument, Not Only a Procurement Step
A contract translates governance questions into enforceable obligations. It must define the
purpose of processing data, whether inputs or outputs are used to improve the service, the

location of processing, retention periods, the deletion mechanism, and the subprocessors.
General phrases such as “we may use data to improve our services” are not sufficient when
the data is sensitive or subject to sectoral restrictions.

The organization also needs a change-management clause. A vendor may change the model,

the model provider, the safety filters, the interface, the price, or the data location. Some of
these changes may affect test results or legal compliance. It should therefore be defined
which changes require notice, how long the notice period is, and what right exists to retest,

refuse, or terminate.

In AI-based services, a service level agreement measuring availability is not enough. A service
may be available yet produce lower quality or change its behavior. You can also ask the

vendor to commit to publishing documentation, to sharing incident logs, and to cooperating
in an investigation. Avoid absolute promises that nobody can measure. What matters most
is that the contract aligns with the internal monitoring and testing plan.

An exit plan should be tested before it is needed. Can the organization export its data,

templates, and logs? Does the process depend on a proprietary format? Are alternative skills
available? What is the expected transition time? Knowing the answer early strengthens
negotiation and prevents a successful project from turning into a dependency that is hard

to control.

Risks in Generative AI Applications
Published risk lists for large language model and generative AI applications set out a set of
risks including prompt injection, disclosure of sensitive information, supply chain risk, data
or model poisoning, insecure handling of outputs, and granting a system excessive executive

capability [5]. The most important message for governance is that textual instructions alone
are not a security control; permissions must be separated, outputs verified, tools restricted,

secrets protected, and complete paths tested.

When dealing with generated or altered images, video, or audio, technical detection alone
is not enough. Provenance, edit history, and content credentials should be preserved where
available, while recognizing that provenance data helps trace the origin and does not

establish the truth of the claim itself [15].

378