Skip to main content
Central IT
Webinar & Cybersecurity

Agentic IT in practice: what the Central IT webinar showed about autonomous agents in cybersecurity and IT operations

In the webinar The Era of Agentic IT (A Era da TI Agêntica), three specialists showed where to start: get the operation in order before putting AI agents on top of it, understand how the attacker thinks, and use agents that prioritize and fix vulnerabilities with a person in control. This is the summary, with the full recording.

By Central IT 10 min read
Full webinar recording, live on October 6, 2026. Recording in Portuguese; the quotes in this article are translations. Watch on YouTube.

What The Era of Agentic IT webinar was

On October 6, 2026, Central IT brought together infrastructure, IT operations, service management, and security professionals to discuss a practical question: how to use autonomous AI agents in IT operations and cybersecurity without creating more risk than they solve. The host was Paullyanna Gomes, Head of Digital Marketing at Central IT, and the program had three talks, from the landscape to practice:

What the registrant survey showed

Before the event, registrants answered the IT Operations and AI Automation Maturity Survey. Most of them work in infrastructure and IT operations (66.8%) and in IT service management or service desk (61.3%). The results presented at the opening show the distance between the discourse and the operation:

Among the biggest IT and security challenges, the pattern was the same: manual tasks, team overload, and a lack of specialized professionals. The bottleneck is still people working at their limit.

To show the scale of the problem, the opening brought a figure from Fortinet's FortiGuard Labs, published by IT Forum: in 2025, Brazil received 753.8 billion attempted cyberattacks. Divided over 365 days, that is about 2.07 billion per day, 86 million per hour, and 23,903 per second.

Emauri Gaspar: AI is not the hero, operational maturity is

Emauri opened with an image: on one side, the stage, where vendors, articles, and influencers talk about AI agents solving everything on their own; on the other, production, with alert fatigue, false positives, configuration errors, and logs nobody can read. In the middle, on a fragile bridge, stands the IT manager.

“AI is not the hero of this story. The hero of this story is operational maturity.”

Emauri Gaspar, at 9:50 in the webinar

The figures he brought, all from Gartner, show how wide that gap is. One in five organizations reports getting value from its AI tools, and one in 50 reports transformational value. And 40% of agentic AI projects will be canceled by 2027. That will not stop adoption: Gartner projects that companies with agentic AI running IT infrastructure will go from less than 5% in 2025 to 70% in 2029. For Emauri, the question is no longer “if” but “when”, because attackers already use AI and the tools companies buy will bring agents anyway. What remains to be seen is whether the operation will be ready for them.

Why do pilots die? Because the bet was placed on the intelligence of the model, not on the foundation of the operation:

“Putting an agent on top of a disorganized operation does not produce autonomy, it will produce a faster incident.”

Emauri Gaspar, at 14:05 in the webinar

The foundation, in his words, is knowing what the infrastructure has, how it works, how services talk to each other, and how each one depends on each asset. It is also basic hygiene, which sophisticated attacks tend to push out of focus: patch management, two-factor authentication, incident response, weak passwords, and outdated servers. It is through this lack of hygiene that most ransomware gets in. The infrastructure manager and the security manager look at the same environment from different angles, and without basic hygiene neither of them succeeds.

The human factor and the AI nobody approved

Emauri's second point was that people remain at the center of risk. The figures he presented come from Verizon's 2025 Data Breach Investigations Report, cited by Gartner, and from a Gartner survey on secure employee behavior:

A document pasted into an AI tool goes to a place the company does not control, and many tools built on top of the big ones do not make clear where the data goes. The proposed answer is not more generic training once a year, but behavior and culture programs, with microtrainings triggered when a real risk shows up in day-to-day work.

Emauri split AI security into two fronts. The AI we use: knowing which third-party tools employees use, the so-called shadow AI, controlling access, and preventing leaks. And the AI we build: the agents that run inside the company's own applications. Even without giving users a prompt field, a system that reads fields filled in by people can receive a prompt injection and carry it all the way to the AI processing. Hence continuous testing and protection against manipulation.

He also pointed to the supply chain: according to Gartner, 59% of CISOs rank attacks through third parties among the top five threats for the next 12 months. Partners, contractors, and suppliers connected to the network, by email or by API, are also part of the attack surface.

Preemptive security and quantum computing

In the Gartner projection Emauri presented for the coming years:

Preemptive security means acting against the attack before it happens, instead of waiting to detect and respond. And the quantum part has already begun: “harvest now, decrypt later” attacks are underway. Encrypted data stolen today can be opened when quantum computing arrives, which means that an attack that will only be discovered in 2029 or 2030 may be happening now.

Three priorities for IT operations in the agentic era

Emauri closed with three priorities:

01Less hype

Evaluate solutions by the results they deliver, not by their market label.

02Foundation

Inventory, CMDB, service maps, basic hygiene, and incident response before any agent.

03Shared governance

Security, technology, and business decide together on risk and AI adoption, with special attention to change management.

On governance, he highlighted change management: every transition or release that adds or changes infrastructure elements can introduce a cybersecurity risk. These are processes that an ITSM organizes, such as change and incident management in CITSmart ITSM. And the subject is no longer only IT's: security, technology, and business need to decide together.

Daniel Donda: how the attacker thinks

Daniel Donda introduced himself as an ethical hacker, someone who studies how adversaries act in order to help companies protect themselves. The first change he pointed out was one of image: today's hacker is not the teenager in the basement, but a profitable criminal business. Just as there is software as a service, there is crime as a service, with groups that sell credentials, tools, and even the attack infrastructure.

These groups do not choose targets by company size, but by opportunity. A small company with one server and five workstations can be hit by ransomware and have neither the team nor the money to recover the environment. Among the cases he sees in incident response are cryptojacking (company servers used by attackers to mine cryptocurrency or host websites) and, at the top, double-extortion ransomware: the environment is encrypted and the data is stolen, with a threat of publication. That is where the LGPD (Brazil's data protection law), the laws of other countries where the company operates, and the damage to its image come in.

And many companies do not even know what they have. Daniel put it this way:

“How are you going to protect what you don't know you have?”

Daniel Donda, at 28:25 in the webinar

AI has sped everything up. Phishing without spelling mistakes in Portuguese has become common, and it is increasingly hard to separate what is real from what is not. That is why, he said, it will become easier and easier to compromise the human being.

The stages of an attack, according to the Cyber Kill Chain

To show how an attack is organized, Daniel used the Cyber Kill Chain, a Lockheed Martin model adapted from a military concept to describe the stages an attacker follows to compromise a system:

01

Reconnaissance

The attacker builds a dossier on the target from what is public: LinkedIn, leaks, technologies in use. With AI, that dossier comes out more complete and faster.

02

Planning

With the dossier, the attacker chooses how to get in. Leaked credentials circulate in criminal marketplaces, so an old password is also a risk.

03

Delivery

Phishing, a fake website, or an unknown USB device. The channel matters less than someone opening the door.

04

Exploitation

The attacker looks for a gap in the environment. Without AI cross-checking the sequence of events, an isolated tool may see each step as normal use of the machine.

05

Installation

The attacker tries to stay in the environment and move from machine to machine, the so-called lateral movement.

06

Command and control

The compromised machine starts communicating with an external server, and data starts to leave.

07

Actions on objectives

Most of the time, the end is ransomware: the environment is encrypted and the stolen data becomes a threat of publication.

None of this happens without leaving traces. The problem is when nobody monitors those traces. Daniel showed a website that tracks ransomware group postings, with cases of 500 GB, 1 TB, and 2 TB of stolen data, taken stealthily. His conclusion: the earlier the chain is broken, the smaller the bill. Stopping the attack at delivery is ideal. Stopping it at installation still requires backup, threat hunting, and forensics, but avoids the final stage.

What the demos teach about defense

Daniel did two live demos (starting at 37:31), with the same lesson: the most common entry point is still a user's decision. A USB device found in the parking lot and plugged into a company computer, or a password that leaked from some website and is still the same as the corporate email password, can be the first step of an attack. With AI, the attacker can also generate likely passwords from what the person exposes on social media.

The recommendations were the basics, done well:

It all starts at the workstation

Vinicius Gravito picked up Daniel's point: the attack started at the workstation. That is where someone clicks the link, downloads the file, or plugs in the device. The server, as he joked, does none of that: it just sits there, calm.

The second problem is that the tools don't talk to each other. Gravito cited Cyber Magazine: a company uses, on average, 30 to 50 applications to do the basics. Infrastructure has its monitoring and inventory tools; security has XDR, antivirus, firewall. The teams talk to each other, but the tools do not. Day to day, an alert comes in from one tool, someone grabs the log, takes it to an AI, looks for context, and meanwhile the attacker has already moved on.

“Today AIs work 24x7. I don't know any security team that works 24x7.”

Vinicius Gravito, at 1:02:33 in the webinar

How CITSmart Autonomous AI works, in the demo

In the demo, Gravito showed CITSmart Autonomous AI, developed by Central IT, running on his own computers. Autonomous works from the edge to the center: an agent installed on workstations and servers, Windows or Linux, handles inventory, ITAM, deployment, and remote access, and collects the machine's telemetry.

The passage that sums up the approach is at 1:04:47 in the recording:

“Autonomous correlates. It is not a SOAR, it is not an XDR, it is not an EDR, it is not a SIEM, but it is all of that too. Why? Because it integrates all these tools using LLM.”

Vinicius Gravito, at 1:04:47 in the webinar

In practice, Autonomous cross-checks what comes from the security and monitoring tools with what is happening on the machine, and cross-checks that with vulnerability databases (CVE, CVSS, KEV, and EPSS) and abuse sources. From there it calculates a risk score and, for each event, provides a summary of what happened, how many times it happened, and an investigation dossier with the confidence level assigned by the LLM. The next steps follow one of two paths: they go to the analyst, with what needs to be done, or to the AI agents, which carry out the action while keeping a person in the loop.

Among what the workstation telemetry showed in the demo:

The indicator dashboard of the demo environment showed the scale of prioritization. There were 22,245 indicators: 7,518 vulnerabilities (CVEs) linked to software installed in the inventory and 14,727 external threat indicators. Of these, 820 were critical (CVSS 9 or higher) and 216 had come in over the last 24 hours. But only 22 vulnerabilities had known real-world exploitation, according to CISA's KEV list, 6 of them used by ransomware. “How many did I have to handle? 22,” Gravito summed up. When the vulnerability is in a software package, the agent deploys the update.

The AI agents work in the automation layer. Each agent has its own skills, the MCP connectors come ready for tools such as Zabbix and for Windows, and credentials are kept in a password vault. Gravito also said that Central IT maintains GPU servers to work with AI models focused on security.

What changes in operations

At the close, Paullyanna Gomes presented the results and benefits that Central IT associates with Autonomous, on four fronts:

She shared the account of a cybersecurity professional in the financial sector: the everyday challenge is the volume of alerts, with no view of how to prioritize. That is the problem correlation and prioritization tackle.

Audience questions

How to request a proof of value

Central IT offers a proof of value for Autonomous: an analysis of the company's environment and technology estate, and the construction of a model according to the needs and routine of each operation. The first step is a conversation with the specialists, which can be scheduled through the link below.

To keep exploring the topic, Emauri Gaspar wrote in TI Inside (in Portuguese) about the agentic era in IT infrastructure management and operations.

Recording chapters

Frequently asked questions

What is agentic IT?

It is the phase in which AI agents stop only answering questions and start acting in IT and security operations: correlating alerts, prioritizing what is critical, and carrying out the fix. In the Central IT webinar, the central idea was that these agents only deliver results on top of an organized operation, with inventory, basic hygiene, and governance, and with a person in control of sensitive actions.

Why are so many agentic AI projects expected to be canceled?

According to Gartner, cited by Emauri Gaspar in the webinar, 40% of agentic AI projects will be canceled by 2027. For Emauri, the reason is that the bet was placed on the intelligence of the model, not on the foundation of the operation: an agent placed on top of a disorganized operation does not produce autonomy, it produces a faster incident.

What is preemptive security?

It is the approach of acting against an attack before it happens, instead of waiting for the attack and then detecting and responding. In the webinar, Emauri Gaspar presented Gartner's projection that, by 2029, products without preemptive capability lose relevance.

What is the Cyber Kill Chain?

It is a model created by Lockheed Martin, adapted from a military concept, that describes the stages an attacker follows to break into and compromise a system: reconnaissance, planning, delivery, exploitation, installation, command and control, and actions on objectives. The earlier the chain is broken, the smaller the damage.

Does CITSmart Autonomous AI replace SIEM, SOAR, XDR, and EDR?

In the demo, Vinicius Gravito explained that Autonomous is not a SOAR, an XDR, an EDR, or a SIEM, but integrates these tools using LLM and cross-checks what they record with workstation telemetry. The idea is to make use of what the company already has; depending on the case, some tools can be integrated or replaced.

Do the AI agents fix things on their own?

It depends on the rule set by the organization. In the demo, the next steps for each event can go to the analyst, with what needs to be done, or to the AI agents, which carry out the fix while keeping a person in the loop. On the product page, Central IT describes three modes: assisted, approved, and automatic.

Where can I watch the webinar recording?

The full recording, in Portuguese, is on the Central IT YouTube channel, with chapters, and embedded on this page. The CITSmart Autonomous AI demo starts at 1:02:28.

Thank you to Emauri Gaspar, Daniel Donda, and Vinicius Gravito, and to everyone who joined live.

Evaluate CITSmart Autonomous AI in your environment

AI agents that correlate what your tools already record, prioritize what is critical, and carry out the fix with a person in control.

Schedule a conversation Explore Autonomous AI