AI Glossary / Governance, ethics and safety / Beginner
Information security
Definition
Information security is the protection of information and the systems that handle it against unauthorized access, disclosure, alteration, disruption, or destruction. Its familiar core objectives are confidentiality, integrity, and availability. For an AI service, it covers the data, models, software, accounts, infrastructure, and operational processes needed to keep information appropriately protected and usable.
What it means
Confidentiality concerns who may see or obtain information. Integrity concerns whether information and system state are protected against improper changes. Availability concerns whether authorized users can access the information and service when needed.
These objectives can differ in importance across assets. A public knowledge article may require little secrecy, but unauthorized changes could mislead many readers. A confidential internal dataset raises a different balance of concerns.
AI does not replace these principles. It introduces additional paths through which information is collected, transformed, retrieved, and presented. A useful security review follows that whole path rather than considering only the model or only the server on which it runs.
How it works
Identify the information and systems involved, who is responsible for them, and what could happen if their confidentiality, integrity, or availability is lost. Map how information enters, moves through, and leaves the service, including suppliers and connected tools.
Choose controls that address relevant risks: identity and access management, appropriate protection of stored and transmitted data, controlled changes, monitoring, backups, and response arrangements. Test whether the controls work as intended, including at boundaries between components.
For an AI assistant, a user may be allowed to ask a question without being allowed to retrieve every connected document. Permissions therefore need enforcement in the underlying system. A natural-language instruction asking the model to keep secrets is not a substitute for an access-control boundary. Review changes and incidents as the service evolves.
Three practical examples
1. A retrieval assistant restricts document access according to the requesting user's permissions, addressing confidentiality throughout retrieval and response generation.
2. A team controls who can change a model's deployment configuration and records approved changes, supporting integrity of the operating service.
3. An organization prepares and tests a fallback for an AI-dependent workflow so that an outage does not leave essential work without a usable alternative.
Think of it this way
Think of a records room with controlled keys, a process for authorized corrections, and arrangements for retrieving files during working hours. A lock supports secrecy, but it does not by itself prevent incorrect edits or ensure that staff can obtain needed records.
The analogy is limited because digital information can be copied and transmitted at scale, and AI can produce new combinations of existing information. The useful point is that security has several objectives. Protecting one does not automatically protect the others.
Worked example
Consider an invented internal assistant with 1,000 documents: 700 available to all staff, 200 restricted to one department, and 100 restricted to a small project team. A particular user is authorized for the first 700 plus 50 project documents, giving 750 authorized documents.
If the retrieval service can search all 1,000 without enforcing that user's permissions, 250 documents lie outside the user's permitted set. That is an exposure in the design, not proof that 250 documents have actually been disclosed.
The team tests that unauthorized documents are excluded from retrieval and checks whether logs, cached responses, and exported results create additional paths. It also verifies that approved documents remain accessible.
This example separates confidentiality from availability and distinguishes a vulnerability from confirmed impact. It shows why a working answer to a question does not establish that the service handled the underlying information securely.
Illustrative case study
Fictional case: Juniper Services connects an AI assistant to internal project records. The initial pilot checks answer quality but does not test users with different permissions. During review, the team discovers that the connector uses a broadly privileged service account without applying user-level restrictions.
The organization limits the connection while redesigning access enforcement and tests representative users before restoring the feature. It also checks whether earlier activity requires incident investigation and whether sensitive information appears in diagnostic logs.
The case illustrates a system-boundary problem, not a claim that every service-account design is insecure. The appropriate architecture depends on how permissions and trust are enforced across the complete workflow.
Common misconceptions and mistakes
Security is not only encryption. Encryption can protect information in particular states, but it does not decide whether an authorized application should disclose it to a particular user.
Privacy and security overlap but are not identical. Secure handling does not automatically establish that every collection or use of personal data is appropriate.
Another mistake is equating model accuracy with security. An accurate answer can disclose confidential information, while a secure system can still make prediction errors. Finally, access granted for one purpose should not be treated as unrestricted permission for every connected AI tool or downstream action.
Business use
Businesses should treat AI information security as part of the existing security program while examining the service's distinctive data flows and dependencies. Inventory connected sources, tool permissions, model assets, logs, and supplier responsibilities.
Prioritize controls according to the consequences of failure rather than applying the same approach to every asset. Include operational staff in recovery and reporting arrangements. Review whether testing covers permissions and change boundaries as well as useful outputs, and repeat relevant checks when a model, connector, or deployment configuration changes.
People also ask
Is information security the same as cybersecurity?
They overlap substantially. Information security emphasizes protecting information, including contexts beyond purely networked digital systems; cybersecurity focuses on cyber systems and their risks.
Does encryption prevent every information leak?
No. Information can still be disclosed through inappropriate access, application behavior, or other authorized-but-misused paths.
Can public information need security?
Yes. Its integrity and availability can matter even when confidentiality is not a significant concern.
Check your understanding
An assistant answers correctly but includes a document the user is not permitted to access. Which core security objective is implicated?
Answer: Confidentiality. Correctness of the content does not make its disclosure authorized. The system's permission enforcement and affected scope need examination.
Related terms
- Integrity
- Availability
- Access Control
- Least privilege
- Incident response
- Confidentiality
- authentication
- authorisation
- data protection
Learn next
Study least privilege and access control next. Trace one AI response from its data source to the final user and identify where permission is checked, where information is stored, and which changes could weaken those boundaries.
More in Governance, ethics and safety
- Individual Bias - Individual Bias is a persistent personal viewpoint, or a narrow set of viewpoints, that influences how someone considers information and makes judgments.
- Individual fairness - Individual fairness is the principle that people who are similar in ways relevant to a particular task should receive similar treatment from a decision system.
- Inference optimization - Inference optimization improves how a trained machine-learning model runs when producing outputs.
- Inherited Bias - Inherited Bias arises when the output of one machine-learning system becomes input for another and carries an existing bias into the downstream learning...
- Input data - Input data is the information an AI system receives to produce a prediction or other inference.
- Insider attack - An insider attack is a harmful action that exploits trusted access or an internal position to misuse an organization's information, systems, or resources.
Sources
NIST CSRC Glossary, information security: https://csrc.nist.gov/glossary/term/information_security
NIST, FIPS 199, security objectives: https://nvlpubs.nist.gov/nistpubs/FIPS/NIST.FIPS.199.pdf
NIST, AI Risk Management Framework 1.0, security and privacy risk discussion: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
Primary definitions and relevant passages were read. The document-access arithmetic and fictional assistant are original examples. Federal categorisation requirements are not presented as universal obligations for private organisations.
Previous: Information retrieval Next: Information security awareness