SecAI+ Domain 3: AI-assisted Security

Darius Cole runs Northstar Financial’s SOC: tens of thousands of alerts a day, a triage queue that never empties. After watching Maya Chen’s team threat-model Nova in Domain 2, Darius pilots not one monolithic “AI platform” but a set of AI-enabled tools placed at specific points in the workflow. Before anything touches production telemetry, he asks Maya the framing question: where does each tool sit, what does it actually do, and what does it expose?

The tool landscape: where the AI lives

SecAI+ names six tool categories. They differ in where they sit, what they can reach, and therefore in what they put at risk. Darius evaluates each one twice: once as a capability, once as an attack surface.

IDE plug-ins

An IDE plug-in embeds an AI model in the integrated development environment, where it sees code as it is written: real-time secure-coding suggestions, completion of detection rules and scripts, and flagging of dangerous constructs — hard-coded credentials, SQL concatenation, weak crypto — before commit. Northstar’s detection engineers use one to draft Sigma rules and SOAR playbook code. The risk runs the other way: most plug-ins send surrounding code to a cloud model as context — secrets, proprietary source, internal hostnames — a quiet data-exfiltration channel if the vendor’s retention and training policies were never reviewed. And AI-suggested code can be subtly insecure or cite packages that do not exist, so it passes the same review gates as human code.

Browser plug-ins

A browser plug-in (extension) brings AI to whatever is on screen: summarizing a vendor advisory, translating a foreign-language phishing page, extracting indicators of compromise from a blog post. Darius’s analysts use one to condense CVE write-ups during triage. The risk profile is defined by extension permissions: one that can “read and change data on all sites” sees webmail, admin consoles, and the SIEM itself — everything the analyst sees, it can potentially transmit. A malicious AI extension is spyware with a helpful UI, so Northstar allowlists vetted extensions and blocks the rest.

CLI plug-ins

A CLI plug-in puts the AI in the terminal. Analysts describe intent in natural language and the tool proposes the command — the exact tcpdump filter, the KQL for a hunt query — or explains command output. It shortens the distance between “I need to” and “I ran it,” which is exactly the danger: an assistant that can execute what it proposes has shell-level authority, and shell history and pasted output routinely carry credentials that travel to the model as context. Darius’s rule: propose freely, execute only with a human confirming each command — Domain 1’s human-in-the-loop pattern.