Agentic Cybersecurity: AI agents for CSIRT, SOC and NOC

Six services, from the first idea to daily operations

They can be hired separately or as a programme. The order follows a typical project: understand, automate, delegate to agents, build to measure, secure and operate.

Assessment and roadmap

We map your security and operations processes, the volume of alerts and tickets, how long each task takes and the tools you already have. With that we prioritise use cases by impact, feasibility and risk, and recommend the right level of autonomy for each one.

A good fit whenYou know there is repetitive work to automate, but not where to start or how much it is worth.

What you get

Process map and indicator baseline
Prioritised use-case matrix
Recommended autonomy level for each case
Roadmap with stages and deliverables

Automation workflows

Playbooks that always do the same thing with the same input: enrich an alert, open a ticket, notify, or block a confirmed indicator. They run on the security orchestrator you already have, or on an open-source base that stays in your hands.

A good fit whenYour team repeats the same tasks by hand every day and the rules for doing them are clear.

What you get

The workflow up and running, with its tests
Documentation and operating manual
Before-and-after measurement

Agentic ecosystems

For work that needs judgement: investigating an alert, building a timeline or drafting a report. An orchestrator agent shares the work out among specialist sub-agents, each with least-privilege tools, controlled memory and human approval gates.

A good fit whenThe tasks need analysis and do not fit a fixed rule, yet they eat up your team’s hours.

What you get

Agents in operation, with their tool catalogue
Approval matrix: what the agent does on its own and what it does not
Quality tests and adversarial tests

Tailored AI solutions

Analyst copilots that answer from your own organisation’s policies, procedures and past cases; report drafters; threat-intelligence summaries; document classification and configuration reviews.

A good fit whenYour team loses time looking for information or drafting documents that always follow the same structure.

What you get

The solution, with its knowledge base
Accuracy tests with real cases
User guide for your team

Agent governance and security

We secure the agents we build and the ones your organisation already uses: a separate identity for each agent, least privilege, a tamper-proof log of its actions and adversarial testing based on the OWASP Top 10 for agentic applications. We review MCP connectors and align everything with ISO/IEC 42001 and with Chile’s personal data law (Law 21,719).

A good fit whenYour organisation already uses AI assistants or agents, or is about to, and needs to know they do not open a new door.

What you get

Assessment of the agents and their connectors
Documented adversarial tests
Controls and usage policies

Operations and continuous improvement

We make sure every workflow and every agent keeps delivering: the quality of its decisions by sampling, false alarms, cost, and changes in your tools or in the models. We fine-tune, add new cases and report the indicators to you regularly.

A good fit whenYou already have automations or agents running and want them to keep performing without burdening your team.

What you get

Sample-based review of the agents’ decisions
Adjustments and new cases as agreed
Regular indicator report for management

What it is for

Narrow, read-only tasks: enriching or summarising.

At any level, anything that cannot be undone or has a high impact always goes through a person.

The scale is inspired by the levels of driving automation (SAE J3016): a simple language for deciding, case by case, how much to delegate.

Agentic Operations: CSIRT · CyberSOC · NOC

The special section of this line. It is where there is the most repetitive work, where deadlines are set by law and where every minute per alert counts. We automate detection, response and operations with agents that work on your tools and under your rules.

How it fits together

An orchestrator receives everything, decides which specialist works on it and respects the autonomy level of each action. People stand at the gates.

SourcesSIEM, EDR, email, identity, cloud, network, service desk, NOC monitoring and reports from people

Orchestrator agentClassifies, shares out the work, brings the results together and respects the autonomy of each action

TriageGroups, removes duplicates and prioritises
EnrichmentIndicators, asset, user and vulnerabilities
InvestigationTimeline, scope and MITRE ATT&CK techniques
Containment and remediationIsolate, block, restart or fail over
CommunicationManagement, customers and status page
ComplianceReporting deadlines, personal data and evidence

Human gatesEvery action that has a high impact, notifies third parties or cannot be undone

Tamper-proof logWhat it saw, what it decided and why, who approved it and what it did

Your own knowledgeYour organisation’s playbooks, runbooks, policies and past cases

Tools are connected with least privilege and with each agent’s own credentials, never a person’s. Every connector is reviewed before it is admitted.

What gets automated in each area

Ten cases per area, with the autonomy level they usually start at and the indicator that measures them. In the assessment we choose with you the ones that pay off most in your operation.

CSIRT · Incident management and response

Intake and classification
Receives reports by email, form or phone, classifies them using your taxonomy and removes duplicates.
Autonomy: N3Measured by: minutes to classify
Case orchestration
Opens the case, assigns it by severity, creates the playbook tasks and tracks internal deadlines.
Autonomy: N3Measured by: tasks completed on time
Law 21,663 clock
Starts the count when an incident is rated as having a significant effect, warns before each deadline and drafts the reports with the case data.
Autonomy: N2Measured by: reports sent on time
Personal data breaches
Helps assess whether notification is required under Law 21,719 and drafts the notices to the authority and to those affected.
Autonomy: N1, N2Measured by: time to decide
Approved containment
Isolates devices, disables accounts, revokes sessions and blocks indicators on firewalls and proxies.
Autonomy: N2, N3Measured by: time to contain
Evidence collection
Triggers evidence acquisition with its hash and chain of custody.
Autonomy: N2Measured by: cases with intact evidence
Communications
Drafts for management, customers, suppliers and staff, in the approved tone and level of detail.
Autonomy: N1Measured by: time to first notice
Third-party coordination
Prepares what is shared with the national or sector CSIRT, the insurer and suppliers.
Autonomy: N1Measured by: response time to third parties
Lessons learned
Builds the post-incident report with the timeline and proposes changes to playbooks and detections.
Autonomy: N1Measured by: improvement actions closed
Exercises
Generates tabletop exercises from real threats and assesses the team’s responses.
Autonomy: N1Measured by: exercises run and gaps found

CyberSOC · Detection and analysis

Alert triage
Groups alerts, removes duplicates, closes known false positives and prioritises the rest.
Autonomy: N3, N1Measured by: alerts resolved without an analyst, and accuracy by sampling
Automatic enrichment
Indicator reputation, asset and user context, and the device’s vulnerabilities.
Autonomy: N4Measured by: minutes until the full context is available
Reported emails
Analyses links and attachments, finds the other recipients and proposes removing the email from every mailbox.
Autonomy: N2Measured by: time per report and emails removed
EDR alert investigation
Process tree, prevalence, timeline and ATT&CK technique.
Autonomy: N1Measured by: investigation time
Identity
Impossible sign-ins, MFA fatigue and privileged accounts; proposes ending sessions.
Autonomy: N2Measured by: time to contain the account
Threat hunting
Proposes hypotheses and queries based on current threats and summarises the findings.
Autonomy: N1Measured by: hypotheses tested and findings
Detection engineering
Drafts rules, tests them against historical data and documents them.
Autonomy: N1Measured by: new rules and their false-positive rate
Prioritised vulnerabilities
Cross-checks known exploitation against the asset’s criticality and exposure, and opens a ticket for its owner.
Autonomy: N3Measured by: days to fix what is critical
Threat intelligence
Summarises public and vendor advisories, extracts indicators and looks for matches in your environment.
Autonomy: N3Measured by: hours between the public advisory and the internal review
Shift handover
Summarises open cases, pending items and risks for the next shift and management.
Autonomy: N4Measured by: handover time

NOC · Network and service operations

Correlation and noise reduction
Groups alarms by common cause using the topology and recent changes.
Autonomy: N3Measured by: alarms per real incident
Assisted root cause
Reviews dependencies, changes and metrics, and proposes the most likely cause with its evidence.
Autonomy: N1Measured by: time to diagnose
Routine remediation
Restarts services, frees up space or fails over to backup following tested runbooks.
Autonomy: N3, N2Measured by: recovery time
Change risk
Before a change is approved, reviews its history, dependencies and window.
Autonomy: N1Measured by: failed changes
Capacity
Anticipates saturation of links, disks and licences.
Autonomy: N1Measured by: capacity incidents avoided
Tickets
Enriches, classifies and routes each ticket to the right team.
Autonomy: N3Measured by: reassignments per ticket
Status communication
Updates the status page and notifies the affected areas.
Autonomy: N2Measured by: minutes to first notice
On-call summary
Summarises what happened during the on-call period for the handover.
Autonomy: N4Measured by: handover time
Post-change verification
Runs automated tests after maintenance or a change.
Autonomy: N3Measured by: faults detected before users notice
NOC–SOC bridge
Recognises when an outage may be an attack and escalates to the SOC or CSIRT with the context.
Autonomy: N3Measured by: time between the outage and the alert to the SOC

N0 manual · N1 assisted · N2 supervised · N3 bounded autonomy · N4 policy-based autonomy. See the autonomy scale

The Law 21,663 clock, automated

When an incident is rated as having a significant effect, the compliance agent starts the count, warns before each deadline and prepares a draft of each report with the data the case already holds. A person reviews it and sends it.

Which deadlines apply to you depends on your organisation's status; we determine it case by case.

3 hoursAlert
72 hoursSecond report (24 hours if an operator of vital importance sees its essential service affected)
7 calendar daysAction plan, for operators of vital importance only
15 daysFinal report, counted from the alert

When an outage may be an attack

The NOC is the first to see that something has stopped working; the SOC knows how to recognise an attack. The NOC–SOC bridge picks up the signs of a denial of service or of servers being encrypted and escalates to the SOC or CSIRT with the context, without waiting for someone to notice.

NOC

A service stops responding

Agent

Recognises an attack pattern

SOC · CSIRT

Opens the case with its context

What we measure and report to you

No case goes live without a baseline and an agreed target. If it does not improve, it is adjusted or switched off.

Mean time to detect (MTTD)
Mean time to acknowledge (MTTA)
Mean time to resolve (MTTR)
Alerts resolved without an analyst
Accuracy by sampling
Analyst hours freed up
Reports to the authority on time
Alarms per incident (NOC)
Failed changes
Actions reversed by a person

How we work

Five stages, each ending with something you can review.

AssessmentProcesses, volumes, times, tools and risks, with the indicator baseline.Ends with the use-case matrix approved by you.
DesignFor each case: inputs, decisions, tools, permissions, autonomy level, approvals and what gets logged.Ends with each case’s specification approved.
Build and testingTested with real past cases and with deliberate attempts to deceive the agent.Ends when it meets the agreed quality target.
Shadow runThe agent works alongside your team without taking any action, and their decisions are compared.Ends when agreement reaches the agreed minimum.
OperationSwitched on at its autonomy level, with sample-based review and regular reports.Continues with improvement: autonomy goes up or down according to the indicators.

Agents you can trust

Giving an agent autonomy opens up new risks. These are the commitments we work by and the controls we apply to every agent.

People in chargeEvery action has its autonomy level declared, and anything irreversible goes through a person.
Measure before and afterNo case goes live without a baseline and an agreed target.
Your data where it belongsModels run on your infrastructure or in an environment we agree with you. An external AI provider only comes in with your authorisation and a data processing agreement.
No vendor lock-inWe work on the tools you already have; if something is missing, we suggest alternatives, including open-source ones.
Fully traceableEvery decision and action an agent takes is logged with its reasoning: useful for the auditor, for the authority and, if it comes to it, for an expert witness.

How we protect the agents

We follow the OWASP Top 10 for agentic applications and the principles of ISO/IEC 42001.

Instructions hidden in what the agent readsExternal content is treated as data, never as an instruction, and is tested with malicious emails and files before going live.
Excessive permissionsLeast-privilege tools, a list of permitted actions per level and human approval for anything high-impact.
Agent identityEach agent has its own short-lived credentials, reviewed regularly.
Supply chainInventory of models, connectors and libraries; connectors reviewed before admission and versions pinned.
Poisoned memory and cascading failuresMemory isolated per case, usage limits and an emergency stop.
Over-reliance on the agentThe agent shows its evidence, not just its conclusion, and we measure how many actions a person reverses.

Part of everything else we do

Agentic Cybersecurity builds on Ziemtinel’s other service lines, and they build on it.

Frequently asked questions

Will AI make decisions on its own in my organisation?

Only the ones you authorise. Every action has an autonomy level agreed in writing, and anything that cannot be undone or has a high impact always goes through a person. Levels are reviewed against the measurements and can be lowered at any time.

Does my data leave my organisation?

No, unless you decide otherwise. By default, models run on your infrastructure or in an environment we agree with you. If an external AI provider makes sense, it is used only with your authorisation and a data processing agreement.

Do I have to buy a new platform?

Not necessarily. We work on the tools you already have and, if something is missing, we suggest alternatives, including open-source ones. The purchasing decision is always yours.

Does this replace my security team?

No. It takes the repetitive work off your team so they can spend their time on what needs judgement: investigating, deciding and improving. People stay in charge.

What happens if an agent gets it wrong?

That is planned for. High-impact actions go through a person, every decision is logged with its evidence, quality is reviewed by sampling and every agent has an emergency stop. If a case does not meet its target, it is adjusted or switched off.

How will I know it is working?

Before we start, we measure your current situation and agree a target for each case. Then we report the indicators to you regularly, compared with that baseline.

Does it help with Law 21,663 or Law 21,719 compliance?

It helps sustain the processes those laws require: classifying incidents, meeting reporting deadlines, preparing notifications and preserving evidence. Decisions and submissions to the authorities remain in your organisation’s hands.

Where is the best place to start?

With the assessment: it shows what to automate first, at what autonomy level and what result to expect. You can also start with a single end-to-end case, to measure it before scaling up.

Do you work with managed security providers?

Yes. If your company provides SOC or incident response services to other organisations, we can design ecosystems that serve several clients, each with its data kept separate.

Cannot find your question? Tell us about your case in the quote tool and we will answer it with the proposal.

Agentic Cybersecurity: agents do the work, people make the calls

Six services, from the first idea to daily operations

Assessment and roadmap

What you get

Automation workflows

What you get

Agentic ecosystems

What you get

Tailored AI solutions

What you get

Agent governance and security

What you get

Operations and continuous improvement

What you get

You decide how much each agent does on its own

N0Manual

The agent

The person

What it is for

N1Assisted

The agent

The person

What it is for

N2Supervised

The agent

The person

What it is for

N3Bounded autonomy

The agent

The person

What it is for

N4Policy-based autonomy

The agent

The person

What it is for

Agentic Operations: CSIRT · CyberSOC · NOC

How it fits together

What gets automated in each area

CSIRT · Incident management and response

CyberSOC · Detection and analysis

NOC · Network and service operations

The Law 21,663 clock, automated

When an outage may be an attack

What we measure and report to you

How we work

Agents you can trust

How we protect the agents

Part of everything else we do

Frequently asked questions

Tell us what you want to automate