AI Security
AI security is more than model safety
AI security includes protecting models and data, and controlling what AI systems are allowed to do in the real world.
Knowledge graph
Authoritative introductions to digital trust. No unverified product claims, no thin keyword pages.
All topicsThis page
AI security
Last reviewed 2026-08-21
Next
Add agent identity and human authorization to the AI threat model.
Request a consultationWho this is for
CISOs, AI platform owners, and public-sector AI governance teams.
Problem
Many AI security guides stop at prompt injection and ignore who the agent is allowed to be.
What to do next
Add agent identity and human authorization to the AI threat model.
Knowledge topics
LibraryAI Security
Definition
AI security is the set of controls that reduce harm from AI systems: confidentiality of training and prompt data, integrity of models, availability of services, and — for agents — identity, authorization, and audit of actions.
AI Security
Why it matters
Generative systems scale impersonation and social engineering. Agents scale action. The second problem is an identity and authorization problem, not only a content-filter problem.
AI Security
Threat model
Prompt injection, data exfiltration, model theft, poisoned tools, deepfake fraud, and agent abuse (an agent using over-privileged credentials) are distinct classes.
Identity architecture does not prevent every class. It is most relevant when an AI system acts on other systems.
AI Security
Standards context
Frameworks such as the NIST AI RMF and the EU AI Act address risk management and legal duties. They are not the same as a technical agent-identity protocol. This site does not give legal advice.
AI Security
Where this meets Keyra
Keyra's public narrative already frames human + AI trust as infrastructure. Detailed agent controls are covered on AI Agent Security and Human Authorization.
Questions
What people ask first.
Plain answers. No product claims that have not been published elsewhere on Keyra.ie.
Does identity stop prompt injection?
No. Prompt injection is an application and model-isolation problem. Identity helps when the model or agent then calls tools and APIs.
Is deepfake detection the same as identity?
Detection estimates whether media is synthetic. Identity asks who is authorized to act. Both can be needed.