Can businesses trust autonomous AI to defend them from cyber threats?

The next significant development in enterprise cybersecurity may not simply be the use of artificial intelligence to analyse threats. It could be allowing AI to take a more active role in deciding what happens next.

Agentic AI represents an important step in that direction. Rather than restricting AI to producing an answer to an individual prompt, the agent concept is centred on systems capable of carrying out more complex workflows towards a defined objective. Threat intelligence teams are dealing with large volumes of information while analyst capacity remains limited. Giving software greater ability to process and act upon this material offers a possible route towards reducing that burden.

Yet new research suggests enterprises are not prepared to exchange their capacity problem for an accountability problem. According to Flashpoint’s Enterprise Threat Intelligence Report: 2026, 25 percent of organisations interviewed are already using agentic AI in production for threat intelligence, with another 20 percent piloting the technology. A further 35 percent are planning adoption within the next 12 to 24 months.  Consequently, 45 percent are either using or piloting agentic AI today, but the findings indicate that adoption is being approached cautiously rather than as a wholesale transfer of responsibility from humans to machines.

AI to solve an AI problem?

Flashpoint finds that AI-related threats are the leading area in which practitioners want to expand intelligence coverage, including deepfakes, AI-accelerated vulnerability exploitation and emerging AI-enabled threat vectors. At the same time, organisations are exploring AI as a means of extending their defensive capacity.  In other words, AI is contributing to the intelligence workload while simultaneously being considered as a tool for managing it.

The operational case for automation becomes clearer from the other findings. Alert noise and false positives were cited by 65 percent of respondents as a leading challenge to threat intelligence effectiveness, while 55 percent pointed to insufficient analyst bandwidth. Fifty percent identified integration gaps across security tools.  An effective agent that filters repetitive information or supports laborious stages within an intelligence workflow could therefore release human specialists for tasks where context and judgement matter more.

A conventional AI assistant that generates an incorrect recommendation creates one type of risk. A more autonomous system able to carry a task further through a workflow introduces a different consideration: what happens when the machine is confidently wrong? Flashpoint says practitioners are evaluating where agents can improve workflows while preserving intelligence quality, suitable controls and human oversight.  

A recurring enterprise AI question concerns whether “human in the loop” controls genuinely provide sufficient protection. As Digital Journal has previously explored, human supervision is often presented as a safeguard when AI influences business processes, yet the effectiveness of the arrangement ultimately depends on whether reviewers possess sufficient information and expertise to challenge what the machine produces.

Cybersecurity makes this problem particularly acute. Threat intelligence can be incomplete, contradictory or rapidly changing. An AI system can potentially save time, but speed ceases to be an advantage if staff then need to spend that saved time checking unreliable outputs.

Hence trust needs to be demonstrated rather than assumed.

Where should the human remain?

The significant issue is not whether cybersecurity organisations should use AI. Flashpoint’s research indicates that many already are. Instead, the question is how much autonomy to provide and under what conditions.

This distinction should prevent agentic AI from being reduced to another enterprise technology fashion. A useful security deployment requires an organisation to determine where machines add sufficient value and where decisions still require human interpretation.

Josh Lefkowitz, Flashpoint’s Co-Founder and CEO, places this issue at the centre of planning for next year. “The opportunity for agentic AI on the defensive side is significant, and our research shows security leaders are ready to invest,” he says. “But they aren’t swayed by the hype. They’re holding agentic AI to a high bar, and they want to know where it improves their work, whether they can trust the output, and where human judgment must remain part of the process.”

This suggests the relevant measure for AI adoption is not how autonomous a security operation can become, but whether increasing autonomy produces a demonstrably better security operation.

Measuring value rather than novelty

This brings the discussion to return on investment. Flashpoint says organisations are looking at metrics including time savings, accuracy, risk reduction and ROI when assessing AI as part of threat intelligence workflows.   Those criteria offer a more sophisticated approach than simply measuring whether a company has adopted the technology.

A useful agent might reduce repetitive investigations or help an analyst identify relevant intelligence faster. Conversely, introducing another disconnected system could exacerbate the very integration difficulties that half of the research respondents already identify. This is why agentic AI and cybersecurity architecture cannot sensibly be considered separately.

There are also financial conditions supporting experimentation. Half of respondents providing a budget outlook expect threat intelligence spending to rise over the coming 12 months and 35 percent anticipate increases of at least 10 percent.  The technology therefore appears to be entering a period in which enterprises have both an operational incentive to test it and, in many cases, additional spending planned.

The transition from an AI assistant to an AI agent represents more than a change in terminology. An assistant helps a person conduct work. Increasing agentic capability asks organisations to consider which parts of that work can be entrusted to software, the conditions on that authority and how human operators can verify that the resulting actions remain appropriate. Flashpoint’s 14 October webinar reflected this shift in emphasis, focusing on AI-enabled threats, investment and adoption, and the future direction of threat intelligence. Agentic AI promises to move artificial intelligence from generating information towards participating more actively in workflows. That creates opportunities for organisations faced with too much data and too little human capacity, but it also increases the importance of accuracy, auditability, control and meaningful human intervention.