A cybersecurity analyst monitors network alerts in a server operations room.

AI Advances Coincide With a Sharp Rise in Cyber Incidents, Marsh Data Shows

Key Takeaways

  • Marsh, the insurance broker, told Reuters that its clients reported 105 cyber incidents in the last two weeks of October 2026, more than double the 48 recorded in the final two weeks of May, a rise of about 119%.
  • Meredith Schnur, who leads Marsh Risk's cyber practice in the United States and Canada, said AI tools are helping attackers identify and exploit longstanding vulnerabilities, including unpatched software, exposed systems, and weak identity controls, at a much faster pace, and that "we could attribute AI to a significant number of these" incidents.
  • Marsh has not established how many of the incidents actually involved AI, and its figures cover only its own clients, not all cyberattacks worldwide, so the data show a coincidence in timing, not proof that AI caused the rise.
  • GlobalData expects global cyber insurance premiums to rise 6.9% to $16.6 billion in 2026, then accelerate to 10% growth in 2027 to reach $18.3 billion, as businesses respond to rising attack frequency.
  • The same week brought a rush of defensive AI news: OpenAI said NATO's core institutions now use its cybersecurity models through the NCIA, and Microsoft launched its MAI Cyber 2 Flash security model at Ignite, so both sides of the cyber contest are now AI-powered.
  • INTERPOL's African Cyberthreat Assessment Report 2026 found AI involved in 55% of reported cybercrime across Africa, with related losses more than doubling since 2024, from $192 million to $484 million.

The Numbers Behind the Headline

The Reuters report, published on 17 November 2026, is short on length and long on significance. Marsh, one of the world's largest insurance brokers, said its clients reported 105 cyber incidents in the last two weeks of October, compared with 48 in the corresponding two weeks of May. That is an increase of 57 incidents, or about 119%, in five months. Put simply, the number more than doubled.

The figures come from Marsh's own client base, which is both their strength and their limit. On the one hand, Marsh insures a large and varied set of companies, so its incident counts are a useful window into what businesses are actually experiencing. On the other hand, the data reflect reports from Marsh clients only, not all cyberattacks worldwide, and Marsh did not release a breakdown by country, industry, customer size, or incident severity. A reader should therefore treat the figures as a signal from one large, well-placed observer, not as a census of global cybercrime.

It is also worth noting what the comparison covers. The two windows are the final two weeks of May and the final two weeks of October 2026. Two-week snapshots can move for reasons that have nothing to do with long-term trends, such as a single large campaign hitting several clients at once, or seasonal patterns in reporting. Marsh itself presented the figures as incident reports from its clients over those periods, and the comparison is striking, but two snapshots do not make a trend line, and the report does not claim otherwise.

What Marsh Says AI Is Doing

The central claim in the Reuters report is not just that incidents rose, but that artificial intelligence is accelerating the pace and scale of cyberattacks. Meredith Schnur, Marsh Risk's US and Canada cyber practice leader, told Reuters in an interview that week that AI is enabling attackers to identify and exploit longstanding vulnerabilities, including unpatched software, exposed systems, and weak identity controls, at a much faster pace than before.

Her most quoted remark went further. "The signs that the attackers are using AI enabled and agentic tools to identify vulnerabilities is happening. So I would say we could attribute AI to a significant number of these," Schnur said. That is a judgement from a practitioner who sees incident reports across many clients, and it carries weight. But it is important to read it exactly as it was said: she attributed AI to "a significant number" of the incidents, not to all of them, and not to a specific percentage.

Schnur also pointed to two forces driving the increase together: the speed and scale of AI tool adoption on the attacker side, and organisations taking longer to contain attacks and respond to incidents on the defender side. That second factor matters, because it means the rise is not only about attackers getting faster. It is also about defenders being slower to close the gap, which is a fixable problem in a way that attacker innovation is not.

The mechanism Marsh describes is not a new class of attack but a faster version of old ones. AI does not need to invent a novel way into a network when there are unpatched systems, exposed services, and weak identity controls already waiting. What AI changes is the speed at which those weaknesses can be found and used, and the scale at which the same techniques can be applied to many targets at once. That framing is consistent with what other reporting has found: AI is being used to automate every stage of the attack lifecycle, from reconnaissance and phishing to extortion and evasion, according to INTERPOL researchers quoted in coverage of its 2026 Africa cybercrime assessment.

Correlation Is Not Causation

The careful reader will notice the gap at the heart of this story, and it is a gap the reporting itself is honest about. The rise in incidents coincides with wider AI adoption, but Marsh has not established how many of the incidents actually involved AI, and its data do not establish that AI caused the 119% rise.

That distinction matters more than it might seem. Several things other than AI could push incident reports up over the same five months. More companies could be detecting and reporting incidents they would previously have missed. Attackers could be running larger campaigns for reasons unrelated to AI, such as geopolitical tension or the discovery of a widely used vulnerability. Insurance clients could be reporting more diligently because their policies or their boards demand it. None of these possibilities is confirmed either, which is the point: the data show two things happening at the same time, and the link between them is a reasoned judgement by Marsh's experts, not a measured fact.

This is not a criticism of Marsh or of Reuters. It is how incident data works. Proving that a specific attack used AI, and then attributing a share of a total rise to that cause, would require detailed forensic analysis of each incident, and that analysis is rarely published. What Marsh has provided is a credible expert reading of its own data, and readers should treat it as exactly that: strong evidence that something is changing, and a plausible explanation for why, but not a proven causal chain.

For businesses, the practical conclusion does not depend on settling the causation question. Whether AI caused the rise or merely accompanied it, the defences that matter are the same: patching software promptly, closing exposed systems, and strengthening identity controls. Those are the very weaknesses Marsh named, and they are the ones organisations can fix regardless of what is driving the numbers.

The Insurance Market Is Already Responding

The second half of the Reuters report turns from incidents to money. According to GlobalData, cyber insurance premiums are set to grow at a faster pace over the next two years as businesses respond to rising attack frequency. GlobalData expects global cyber insurance premiums to rise 6.9% to $16.6 billion in 2026, before accelerating to 10% growth in 2027 to reach $18.3 billion.

Those figures describe the insurance market, not the cost of attacks themselves, but they are a useful measure of how seriously businesses are taking the threat. When premiums rise and more companies buy cover, it usually means boards have accepted that incidents are more likely and more expensive than they assumed. It also means insurers are pricing in the same risk Marsh is describing, which is a form of independent confirmation: the people who pay when attacks succeed are charging more.

The growth in premiums also raises a question for smaller businesses. If cover gets more expensive, some companies may go without it, which would leave them carrying the full cost of an incident themselves. Insurers, for their part, are increasingly asking for evidence of basic defences, such as multi-factor authentication and patching discipline, before they will write a policy at all. The Marsh data, in that sense, is not just a news story. It is a preview of the conditions businesses will face when they next renew their cover.

The Same Week, on the Defensive Side

The Marsh figures landed in a week when the defensive side of AI security also made headlines, and the contrast is instructive. On 17 November 2026, OpenAI said it is working with NATO's Communications and Information Agency, the NCIA, to give NATO's core institutions access to its advanced cybersecurity models through its Daybreak programme, with NATO having used up to 300 billion raw input and output tokens a month. The same OpenAI post said the EU's cyber agency, ENISA, used the models to identify four vulnerabilities in software used across EU institutions, with fixes published for all four, and that CERT Polska used OpenAI models to identify six vulnerabilities in MikroTik RouterOS, including two rated CVSS 9.2. OpenAI also supported NATO's Crossed Swords exercise, held in Tallinn from 2 to 6 November, where participants built and tested AI agents for cyber operations.

On the same day, at its Ignite conference, Microsoft announced MAI Cyber 2 Flash, a specialist security model built for continuous cybersecurity monitoring and defensive work, including identifying and patching vulnerabilities in live production systems. Microsoft says the model matches the quality of OpenAI's GPT-6 Astra on the CyberGym-E2E benchmark at less than 2% of the cost per task, and reports a score of 92.3% against 91.5% for GPT-6 Astra on that benchmark. Those are Microsoft's own reported figures, not independently confirmed results, and access is limited to verified defenders inside Microsoft's Codename MDASH programme, with wider availability in Microsoft security products from mid-December. The pattern, though, is clear: the same week that Marsh reported attackers moving faster with AI, the two largest Western AI companies were announcing AI tools for defenders.

That symmetry is the real story of this moment in cybersecurity. AI is not a weapon that belongs to one side. It is a general capability that speeds up whichever side uses it better. Attackers are using it to find weaknesses faster. Defenders are using it to find the same weaknesses first. The contest now turns on who deploys it more effectively, and on the unglamorous basics: whether defenders patch, and how quickly.

The Picture From Africa

The Marsh data is not the only evidence pointing in the same direction. INTERPOL's African Cyberthreat Assessment Report 2026, drawing on survey data from 36 African member countries, found that artificial intelligence is enabling 55% of reported cybercrimes across Africa, making attacks faster, more scalable, and harder for victims and platforms to detect. The report also found that cybercrime-related losses have more than doubled since 2024, rising from $192 million to $484 million, driven primarily by AI-facilitated scams, credential harvesting, and automated social engineering campaigns. Ransomware accounted for 28% of all cyberattacks in Africa in 2025, making it the single largest type of attack, according to the report.

The Africa figures come with the same caveats as the Marsh data, and arguably larger ones. The 55% figure is based on survey data from member countries, and the threshold for counting an incident as AI-enabled is not defined in the public summaries. But the direction is consistent with the Marsh numbers, and the scale of the losses, more than doubling in two years, is hard to ignore. Africa's experience also underlines a point that applies everywhere: with more than 1.1 billion mobile subscribers recorded in 2025, digital adoption is spreading faster than regulation and digital literacy can keep up, which widens the attack surface every year.

Japan, and the Investigations Already Under Way

The rise in incidents is also visible in the work of investigators. On 15 November 2026, we reported on a Japan-led probe into AI-driven cyberattacks affecting more than 1,000 servers, a reminder that the shift Marsh describes is already the subject of international law-enforcement attention. Cross-border investigations of this kind are slow, and they rarely produce quick public results, but their existence shows that governments are treating AI-enabled attacks as a serious and growing problem rather than a theoretical one.

For organisations, the lesson of the Japan probe and the Marsh data together is that AI-enabled attacks are not a future risk. They are a current one, and they are being investigated, insured against, and defended against in parallel. The gap between those three activities, investigation, insurance, and defence, is where most of the risk sits.

What Businesses Should Do Now

The Marsh data points to a short list of practical steps, and none of them requires buying a new AI product. The vulnerabilities Marsh named are familiar ones, and the fixes are familiar too.

First, patch promptly. Unpatched software was the first weakness Schnur named, and it remains the easiest way into most networks. Organisations that know their patching backlog should treat it as their highest priority, because attackers are now finding unpatched systems faster than before.

Second, close exposed systems. Internet-facing services that do not need to be public should not be public, and those that must be public should be monitored. Exposure is a choice, and it is one that AI-assisted scanning punishes more quickly than it used to.

Third, strengthen identity controls. Weak identity management was the third weakness named, and it covers everything from passwords without multi-factor authentication to over-privileged accounts and stale credentials. Identity is now the front door for most attacks, and it deserves the same attention as the network perimeter once received.

Fourth, shorten response times. Schnur identified slow containment and response as one of the main factors driving the increase, which means that detection and incident response are as important as prevention. An organisation that finds an intrusion in hours rather than weeks will suffer far less, even if it cannot stop every attack.

For readers who want a structured starting point, our AI agent security checklist covers the controls that matter most when AI agents and tools are part of the environment, and the UK government's £600 million AI security funding shows how seriously states are now treating this as a national priority rather than a private IT problem.

What We Could Not Verify

  • How many of the 105 incidents actually involved AI. Marsh has not established this figure, and no breakdown has been published.
  • The breakdown of the incidents by country, industry, customer size, or severity. Marsh did not release this information.
  • Whether the two-week comparison windows are representative of the wider trend. The figures cover the final two weeks of May and the final two weeks of October 2026 only.
  • The precise definition of "AI-enabled" used in INTERPOL's 55% figure for Africa. The public summaries do not define the threshold, and the underlying survey methodology has not been published in full.
  • Microsoft's benchmark claims for MAI Cyber 2 Flash. The CyberGym-E2E figures are Microsoft's own reported results and have not been independently reproduced.

Sources

Similar Posts