AI Security
What this domain tests
AI Security covers two fronts: protecting enterprise use of AI applications and protecting AI systems themselves. Expect shadow AI discovery, data loss prevention in prompts, prompt injection, model and dataset posture, and runtime controls.
Core rules
- AI access security discovers sanctioned and unsanctioned generative AI apps, applies acceptable-use policy, restricts tenants, and inspects prompts and uploads. Blocking a domain is weaker than classifying the app and controlling the data flow.
- AI runtime security sits inline with model APIs and applications. It looks for prompt injection, jailbreak attempts, sensitive data in prompts, unsafe model output, and abuse. Keyword filters alone miss paraphrased or encoded attacks.
- AI security posture management inventories models, datasets, plugins, and permissions, then finds excessive access, unencrypted stores, and vulnerable dependencies.
- Governance needs data classification, approved tools, logging, red teaming, and a response plan for leaks.
Worked example
An engineer pastes source code into an unsanctioned chatbot. AI access security identifies the app, DLP recognizes code and credentials in the prompt, and policy blocks the upload while coaching the user toward an approved enterprise assistant. The reason is data flow: the risk is not the chatbot domain alone but sensitive content leaving the organization. The approved tool can log usage, apply retention controls, and restrict training on company data.
Traps that cost marks
Do not rely only on URL blocklists; users can reach AI through APIs and embedded apps. Do not assume a prompt filter solves prompt injection. Do not treat an AI app as ordinary SaaS without checking data retention and model training terms. Do not ignore over-permissioned plugins or datasets. Do not deploy AI monitoring without a logging and escalation plan.