Insights

Splunk .conf26: AI security triage needs evidence and human approval

10 November 2026

AI security triage can prepare evidence without taking control. Learn what Splunk .conf26 means for alerts, escalation and human-approved response.

Cisco’s account of Splunk .conf26 describes a live security operations centre using Splunk to bring security data together and help analysts review the AI Triage Agent’s work. For Australian businesses considering AI security triage, the important distinction is between preparing an investigation and authorising a response.

In its 8 October 2026 account of the event, Cisco describes analysts validating agent-generated findings against evidence. That is a useful operating model, not proof that every business should hand security decisions to an autonomous system.

What Cisco demonstrated at Splunk .conf26

According to Cisco, Splunk Enterprise Security provided the shared investigation platform. Data sources included identity activity, endpoint events, firewall records, network addressing and name resolution information, packet-level evidence and exposure context.

The team also tracked agent activity, including searches, evidence reviewed, confidence assessments and proposed outcomes. This made the agent’s contribution visible alongside the investigation itself.

The setting matters. Conference networks contain demonstrations, unfamiliar devices and unusual traffic that may be legitimate. Cisco’s account illustrates why an alert needs context before someone decides whether it represents a threat.

For a smaller business, the lesson is not to reproduce an event SOC. It is to connect the records needed for a decision and make both the agent’s reasoning and the analyst’s approval traceable. The reported event activity does not establish likely accuracy, savings or response times for your environment.

AI security triage is not autonomous remediation

A useful scope separates four types of work:

| Stage | What the agent can prepare | Required control | |---|---|---| | Alert enrichment | Add device ownership, account role, related events and asset importance | Use authorised sources and show when information was collected | | Evidence summary | Assemble a timeline, reference supporting records and identify gaps | Keep observed facts separate from interpretations | | Escalation | Suggest urgency, draft a ticket and route it to a designated reviewer | Apply documented escalation rules and confirm someone owns the case | | Response | Recommend session revocation, account restriction or endpoint isolation | Require an authorised person to approve the action before execution |

The first three stages help a person investigate. The fourth changes the business environment and can interrupt legitimate work.

Even an apparently minor action can have consequences. Restricting an account may stop a payroll run; isolating a shared workstation may disrupt dispatch. Human approval must therefore be a technical control, not simply a sentence in a procedure.

Ask whether the triage identity can execute response commands directly. If it can, the boundary between investigation and remediation needs closer examination.

Start with reliable evidence, not a polished summary

AI security triage depends on the records it can access. A fluent explanation cannot compensate for missing sign-in logs or an outdated asset register.

Before a pilot, check:

  • Coverage: Which identity, endpoint, email and network records are available, and which are absent?
  • Timing: Are timestamps consistent, and how late can events arrive?
  • Ownership: Can an address or device identifier be tied to the right person at the time of the event?
  • Provenance: Can a reviewer open the original record behind each significant claim?
  • Retention: Will supporting evidence remain available after the initial investigation?
  • Failure handling: Does a failed query appear as missing evidence rather than a clean result?

For example, an overseas sign-in should not become a confirmed compromise solely because its location is unusual. An investigator may need device details, authentication results, session activity and the user’s travel context.

Our Microsoft 365 impossible-travel guide explains why corroborating evidence matters. The same principle applies when an agent prepares the investigation.

Make summaries and escalations usable

A useful AI security triage summary should let the reviewer understand the issue without hiding the uncertainty. Require a consistent format:

  • What triggered the investigation and when.
  • Which accounts, devices and business services are involved.
  • What the underlying records establish.
  • What remains uncertain or could have a legitimate explanation.
  • What further checks or actions are recommended, and why.

Confidence is supporting information, not permission to act. Ask what the confidence assessment represents and whether reviewers can see contradictory evidence.

Escalation also needs an owner, a priority and an acknowledgement path. A draft ticket in an unattended queue is not an effective handover. Define who receives urgent cases, who covers absences and what happens outside business hours.

Avoid allowing the agent to silently close findings because it considers them harmless. Start with proposed outcomes for human review, and record why reviewers accept or reject them. For broader investigation scope, use our threat-hunting checklist.

Keep permissions and response approval separate

Tech Engine agents operate in a private Australian-hosted environment and keep human approval for response actions. That deployment model should be supported by clear access restrictions, evidence handling and approval records.

For any proposed security workflow, verify these controls:

  • Give the triage agent a dedicated identity with the minimum access required.
  • Separate investigation permissions from permissions that modify accounts, devices or network rules.
  • Present the exact target, action, reason and likely operational impact to the approver.
  • Tie approval to that specific action rather than granting open-ended authority.
  • Record the approver, time, executed action and result.
  • Provide a way to stop the workflow and recover from an inappropriate change where possible.

Private hosting alone does not settle every data-handling question. Check where model processing, logs, backups and support access occur, and whether any external enrichment service receives sensitive information.

Security records can also contain attacker-controlled text, including email content and URLs. Treat that material as evidence, not instructions. It must not be able to change the agent’s permissions or bypass approval. Our AI agent identity security guide covers the separation of identities, access and approvals.

Pilot a narrow workflow and measure review quality

Begin AI security triage with one repeatable investigation, such as assembling evidence for suspicious sign-ins. Use read-only access, a defined review group and an agreed escalation procedure. Do not begin by granting broad remediation access.

Test known benign activity, suspicious activity, incomplete records and unavailable connectors. Include misleading text inside sample evidence to check that the workflow respects its instructions and tool boundaries.

Measure whether summaries reference valid evidence, whether important facts are omitted, how often analysts correct conclusions and whether escalations reach the right person. Track investigation time alongside these quality measures. Query volume alone does not tell you whether an investigation improved.

Set stop conditions before deployment. Unsupported claims, exposed sensitive data, unapproved tool calls or attempted approval bypasses should pause the pilot for review.

Review your security workflow before adding autonomy

Cisco’s event account offers a practical lesson: agents can assemble investigative material while people retain responsibility for consequential decisions. Better preparation should make human review clearer, not remove it.

Tech Engine can help review your cyber security and assess managed AI agents that integrate with existing systems, operating privately in Australia with human approval for response actions. Call 1300 088 324, email sales@techengine.au or visit aiaas.au to discuss an appropriately scoped workflow.

Want this applied to your business?

Request an AI Blueprint and we will map the processes worth automating first.

Request an AI Blueprint