What is AI security?
AI security is the work of protecting AI systems from attack and misuse. It covers models, training and retrieval data, prompts, agents and their tools, and the servers that run them. It also covers the reverse risk, that a manipulated AI system becomes a route into the rest of the business. The same phrase is sometimes used for applying AI to security operations, which is a separate topic.
AI systems need every ordinary control, including patching, network segmentation, encryption, and access management. They also add a new problem: a language model cannot reliably tell instructions from data. Any email, web page, or document it reads can carry instructions that change its behavior. When the model can call tools, those instructions can turn into actions.
How AI security works
Many teams map AI-specific threats to controls using the OWASP Top 10 for Large Language Model Applications.
| Threat | What it looks like | Main control |
|---|---|---|
| Prompt injection | Hidden instructions in a message or document redirect the model | Treat retrieved content as untrusted and limit what the model can do |
| Data leakage and exfiltration | The assistant reveals files the user may not open or sends data out | Permission-aware retrieval and blocked outbound traffic |
| Data poisoning | Tainted training or reference data skews answers | Data lineage and controlled ingestion |
| Model supply chain | A downloaded model file runs hidden code when loaded | Verified sources and file formats that cannot execute code |
| Excessive agency | An agent with broad permissions acts on a manipulated request | Least privilege and human approval for changes |
AI security in practice
A hospital network deploys an assistant that drafts replies to patient portal messages. Clinicians sign in through single sign-on, and the assistant reads only the chart of the patient who sent the message. Incoming messages are treated as untrusted input, because anyone can type instructions into one. The model runs on the hospital’s own servers with no outbound internet access, and a clinician approves every reply. Before launch, a red team tries to make the assistant reveal other patients’ records.
Kirk Tech Solutions typically reviews an AI deployment against the client’s existing security standards first, then adds retrieval permissions, tool scopes, and an approval step for any agent action that changes a system of record.
What to watch for
- Check that the retrieval index enforces source-system permissions at query time, because an index built under one service account exposes everything it holds.
- Validate model output before it reaches a database, a web page, or a command line.
- Log every prompt, retrieved document, and tool call, so that an incident can be traced afterward.
- Repeat red-team tests after any change to the model, prompts, or tools.