AI Glossary / Governance, ethics and safety / Beginner
Cybersecurity
Definition
Cybersecurity is the management of risks to digital systems, services and information from unauthorized access, disruption, misuse or other harmful activity. It combines technical protections, organizational decisions and operational practices to preserve trustworthy digital operations.
What it means
For an AI service, cybersecurity includes the surrounding application, accounts, data stores, networks, integrations and people. A well-performing model can still sit inside an insecure service. Conversely, a secure service can produce inaccurate model outputs; security and predictive quality are related but distinct concerns.
Three familiar objectives are confidentiality, integrity and availability. Confidentiality concerns inappropriate disclosure, integrity concerns unauthorized or improper change, and availability concerns timely access for authorized use. Different systems place different demands on each.
AI adds particular assets and failure paths, such as model files, training pipelines and tools an agent can call. It also inherits ordinary security problems, including stolen credentials and incorrectly configured permissions.
How it works
Identify what needs protection and how the system operates. Map important dependencies, data flows, authorized roles and plausible threat scenarios. Choose controls that address those scenarios and maintain evidence that they work.
NIST's Cybersecurity Framework 2.0 organizes outcomes under Govern, Identify, Protect, Detect, Respond and Recover. These functions help coordinate responsibilities rather than prescribing one product or a rigid sequence. Governance guides decisions about priorities and accountability; operational activities protect systems and prepare for failures.
For AI, examine both model-related and conventional application paths. A harmful instruction embedded in retrieved content is a different issue from an exposed integration credential, although both may lead to unauthorized actions. Detection, response and recovery should cover the complete service, including external dependencies and the consequences of partial actions.
Three practical examples
- A company gives an AI assistant narrowly scoped access to business records instead of sharing an unrestricted administrator account.
- A model deployment pipeline checks the provenance and integrity of approved artifacts before serving them.
- A service team practices restoring an AI application's required data and configuration after a simulated operational disruption.
Think of it this way
Think of protecting a library's operations. Locks help control entry, catalog records help preserve order, staff notice problems and recovery plans help reopen after damage. One strong lock cannot address every way the service might fail.
The analogy illustrates multiple objectives and responsibilities. Digital systems have distinct technical mechanisms, so actual security decisions require knowledge of the application's architecture and threats.
Worked example
Access scope changes the possible consequence
An invented assistant needs to read a collection of 500 public product records. Its integration account mistakenly has read and write access to 50,000 internal records.
Changing the account to read only the approved product collection reduces both the accessible record scope and the available operations. The number of accessible records falls from 50,000 to 500, but the team should not translate that ratio into a claim that security risk fell by 99%.
Risk depends on record sensitivity, failure likelihood and other paths as well as counts. The important verified change is that attempts to read outside the collection or write records are rejected by the service. The team tests those restrictions and separately checks whether the assistant can still perform its intended task.
Illustrative case study
Fictional company Maple Research deploys an AI document assistant with strong login controls. During review, staff discover that the document index contains passages copied from restricted folders without preserving their access restrictions.
The team corrects the retrieval authorization path and investigates which information could have been returned. It also updates the ingestion process so that future indexing respects the same access model. Login remains necessary, but it was not sufficient to protect each document.
The company treats the issue as a service-wide security problem involving data flow and permissions, rather than assuming that changing the model or adding an instruction would resolve it.
Common misconceptions and mistakes
- Cybersecurity is only an IT department's concern. Product decisions, purchasing and operational practices can change exposure.
- A model with good accuracy is secure. Accuracy does not establish access control or resistance to misuse.
- Buying a security tool completes the work. Configuration, maintenance and response still matter.
- Preventing every incident is the only objective. Limiting consequences and recovering effectively are also essential.
Business use
Businesses use cybersecurity to protect their ability to operate and the information entrusted to them. AI projects should join the organization's security processes rather than create an isolated exception because the technology is new.
Ask who owns each integration, which data it can access, how changes are reviewed and what happens when a credential or dependency is compromised. Evidence from permission tests, monitored operation and recovery exercises is more useful than broad claims that a model is secure. Review the arrangement when permissions, data sources or tools change.
People also ask
Is cybersecurity the same as privacy?
No. Security protects systems and information; privacy also concerns how information about people is collected and used, including authorized uses.
Does encryption solve all security problems?
No. It can protect data in particular states or channels while leaving other weaknesses untouched.
Can AI help cybersecurity?
Yes, but AI-assisted detection or analysis still requires suitable evaluation, oversight and protection of its own workflow.
Check your understanding
An AI service requires login but returns documents that the logged-in user is not permitted to read. Is authentication alone sufficient?
Answer: No. The service also needs appropriate authorization for the requested information. Knowing who the user is does not establish what they may access.
Related terms
Learn next
Read access control and incident response. Trace one AI request from user login through retrieval and tool execution. Identify where each permission is enforced and how an incorrect action would be detected and addressed.
More in Governance, ethics and safety
- Coverage bias - Coverage bias occurs when the population represented in a dataset does not adequately match the population about which conclusions or predictions are intended.
- Criterion validity - Criterion validity concerns how well a measurement relates to an appropriate external reference or outcome.
- Crowdsource - To crowdsource is to invite contributions from a distributed group of people to complete a task, supply information or solve a problem.
- Dark pattern - A dark pattern is a design practice that misleads, pressures or obstructs people in ways that undermine informed choice.
- Data control - Data control is the practical ability to set and enforce conditions on how data is collected, accessed, changed, used, shared and retained.
- Data dredging - Data dredging is searching through many analyses or patterns and presenting selected favorable findings without adequately accounting for the search.
Sources
NIST, The Cybersecurity Framework 2.0: https://nvlpubs.nist.gov/nistpubs/CSWP/NIST.CSWP.29.pdf. NIST's official framework overview explains its six functions: https://www.nist.gov/news-events/news/2024/02/nist-releases-version-20-landmark-cybersecurity-framework. The examples are original explanations of security principles, not a complete assessment or guarantee for a particular deployment.
Previous: Cyber kill chain Next: Cycle