Security operations teams have more telemetry than people to read it. Endpoint agents, identity providers, cloud logs, email gateways and network sensors all produce alerts, and most of them are noise.
AI can help. It can group related alerts, summarise an incident, suggest the next query and draft a report. It can also make confident mistakes, and an AI system with broad access is itself an attractive target.
That is why AI in the SOC and zero-trust architecture belong in the same conversation.
What AI is good at in security operations
From building and running production systems, the useful jobs are specific:
- Triage — clustering related alerts and ranking them by likely impact
- Enrichment — pulling asset owner, recent changes, identity context and threat intelligence into one view
- Investigation support — proposing queries, explaining unfamiliar log formats, building timelines
- Writing — incident summaries, handover notes and first drafts of post-incident reviews
- Detection engineering — suggesting rule changes from past false positives, reviewed by an analyst
Notice what is missing: unattended containment of production systems based only on a model’s judgement. Automated response has a place, but it should run through playbooks with explicit rules, not open-ended model decisions.
Zero trust in plain terms
Zero trust means no user, device or service is trusted because of where it sits on the network. Every request is checked against identity, device health and policy, and access is limited to what the task needs.
In practice that looks like:
- Strong identity for people and for workloads
- Least-privilege access, granted per task and time-bound where possible
- Segmented networks so one compromise does not reach everything
- Continuous verification rather than a single login at the start of the day
- Logging of every access decision
Standards bodies such as NIST publish reference architectures for this. The principles are more important than any single product.
Treat the AI assistant as a privileged identity
An AI tool that can read logs, query endpoints and open tickets is a service account with a lot of reach. Apply zero-trust rules to it like any other identity.
- Give it its own workload identity, not a shared analyst credential
- Scope read access to the data sources it needs
- Keep write actions, such as isolating a host or disabling an account, behind approval
- Log every tool call with the request, the parameters and the analyst who asked
- Review its permissions on the same cadence as human administrators
Prompt injection is a real risk here. Attackers control some of the text your SOC reads: email bodies, file names, web pages, user agent strings. If a model processes that text and can also take actions, hostile content can try to steer it. Separate untrusted input from instructions, and never let model output trigger a privileged action without a policy check.
Keep analysts in charge
AI should shorten the path to a decision, not replace the person making it.
Good design shows the evidence behind each suggestion, links to the underlying events and makes it easy to disagree. When an analyst overrides a recommendation, record why. Those overrides become test cases and tuning data.
Measure what matters to the team: time to triage, time to contain, false positive rate and analyst workload. If the tool adds a review step without removing work elsewhere, it is not helping.
Data handling in the SOC
Security logs contain personal data, credentials in error messages and sensitive business detail. Decide early where models run, what data leaves your environment and how long prompts and outputs are retained. Where policy requires it, use private or on-premise deployment.
Where to start
- Pick one high-volume, low-risk workflow, such as phishing triage or alert enrichment
- Give the AI component its own identity with read-only access at first
- Build an evaluation set from past incidents with known outcomes
- Log every model call and tool action in the same place as other security events
- Add write actions one class at a time, each behind an approval gate
Incubics builds AI tooling for security teams with controls designed to support zero-trust and audit requirements from the first release.