For more than two decades, Cybersecurity Awareness Month has focused squarely on people: the employee who might click, the finance manager who might wire money to the wrong account, the contractor who reuses a password. This October, the people responsible for enterprise defences are pointing at a different workforce, one that never sits through a training session and keeps working long after everyone has logged off.
Agents now outnumber employees
Edwin Weijdema, Field CTO EMEA and Strategy & Cybersecurity Lead at Veeam Software, puts a number on it. Non-human identities- the service accounts, API keys and AI agents operating inside corporate systems- now outnumber human employees by 109 to one.
A year ago, Weijdema argued that organisations had moved beyond awareness and needed to concentrate on preparedness. He has since revised that view. “This year, I believe organisations need to shift back to awareness, so they can understand just how fundamentally AI has changed the structure of their digital estate.”
His real target is how companies measure their security. Compliance status and periodic assessments capture a single moment, and “in an environment where overprivileged AI agents are constantly accessing data and acting at machine speed, these metrics are outdated from the very second they are measured.” The standard he proposes is considerably tougher. It is no longer enough to be compliant or protected, he says; organisations must be able to prove it, “not just at predetermined yearly or quarterly assessments, but at the very second that an agent makes an error.”
How much should AI decide on its own?
That shift is already reaching the boardroom. Alexandre Depret-Bixio, SVP International at Anomali, says AI’s move from advising humans to acting on their behalf has left every board member and C-suite leader with the same question: what do we allow machines to decide? Security leaders want what autonomy promises, including threat hunting that carries on while their teams sleep and policy-driven decisions applied in real time across their systems. What holds them back, in his view, is less the intelligence of the models than the quality of the data feeding them. “If you can’t trust your data, you can’t trust the decisions AI agents make on your behalf,” he says. “You end up with confident but bad decisions, and in cybersecurity, that risk is untenable.”
Depret-Bixio’s answer is a gradual one. “Autonomy earns trust one governed decision at a time, with human approval and audit where it matters.” Across the GCC, where AI adoption and digital transformation are moving quickly, that balance between automation and oversight is turning into a practical decision for security teams rather than a question for a future strategy paper. “The near-term future of AI will be judged by whether we can stand behind every decision it makes,” he says.
Controlling what AI agents can reach
Frank Ruge, SVP EMEA at Infoblox, comes at the problem from the network. “AI safety is not only a model problem,” he says. “Agents are systems with access to data, networks and services. The control question is simple: what can they reach, what access do they have, who owns them, and how do you stop them?” Infoblox argues that DNS offers an early point at which IT teams can see and control what an agent connects to, while acknowledging that it is one layer among several.
Containing AI also depends on identity controls, segmentation, egress restrictions, telemetry and response workflows that have actually been tested. Ruge sees particular urgency in the UAE, which is placing AI into high-value public and private sector services. “Every autonomous agent needs clear boundaries around the services it can discover and contact, with policy and visibility at the network layer, not assumptions that the application boundary alone will hold.”
None of this replaces the fundamentals. Ray Heffer, Field CISO at Veeam, wants the month to act as “a catalyst for turning security fundamentals into operational discipline,” which means strong authentication, phishing awareness, timely patching, least-privilege access and recovery plans that have been tested. AI widens the scope of that discipline, he says, to include “clear visibility into where data lives, how it flows, who or what can access it, including AI agents, and whether clean, trusted data can be recovered when something goes wrong.” Done properly, resilience is what lets a business “say ‘yes’ safely” to new technology instead of stalling it.
AI is removing the usual signs of phishing
The human target has not gone away either, although the old advice on spotting it has aged badly. Derek Manky, Chief Security Strategist and Global VP Threat Intelligence at FortiGuard Labs, notes that the spelling mistakes and clumsy phrasing that awareness programmes once trained people to look for are no longer reliable warning signs. AI-generated messages can be polished, personalised and written in the familiar style of a colleague or supplier. Spear phishing used to require an attacker to research a target’s role and relationships before writing a convincing message, and AI now shortens each of those steps, while synthetic images, audio and video make even a recognisable voice weaker evidence of who is really on the line.
Attackers also follow the news. FortiGuard Labs counted more than 13,000 domains linked to the FIFA World Cup 2026 registered between January and May, with around 8.8% judged malicious or suspicious, offering fake tickets, job adverts, streaming links and impersonated accounts. Manky’s argument is that defenders should care less about how a message looks and more about what it is asking for. “The more useful question is no longer, ‘Does this message look professional?’” he says. “It is, ‘Does this request make sense, and have I verified who is making it?’” That means checking any request for credentials, payments, MFA approvals or urgent action through a separate, trusted channel, and following approval procedures even when the request appears to come from an executive.
For banks and fintechs, AI-powered fraud is now an enterprise problem as much as a consumer one. Manish Mimani, Founder and CEO of Protectt.ai, says the digital ambitions of the UAE and Saudi Arabia depend on people trusting the services they use every day, and that attackers are now generating “hyper-realistic identities, convincing messages, and sophisticated scams at unprecedented scale.” No single bank, fintech or security firm can absorb that alone, he argues, and he calls for closer collaboration across BFSI and cybersecurity players to share threat intelligence and deploy defences that keep pace.
Breach reporting deadlines are getting shorter
When defences do fail, the clock now starts sooner and runs faster. James Maude, Field CTO at BeyondTrust, says the disclosure timeline “used to start when the legal team said it did. Now it can start when the SOC analyst responds to an alert at 2:00am.” Most incident response plans, he says, were not designed around GDPR’s 72-hour window or the SEC’s four-business-day materiality filing. The first hours have become a race to answer one question: what did the attacker actually have access to? “If you cannot map the blast radius of compromised identities, you cannot scope the breach, and you certainly cannot disclose it with any confidence.”
Santiago Pontiroli, Lead TRU Researcher at Acronis, says companies now have to investigate, contain and prepare notifications at the same time, beginning with a reliable incident record of what is confirmed, which systems and identities are affected, and who owns each decision. He picks out one point that technical teams often overlook. Removing an infostealer from a device does not invalidate the credentials or sessions it has already stolen, which may still give an attacker a way back in. For multinationals, Maude suggests designing around the strictest regime. “If you can meet a 24-hour notification in one region, you can meet 72 hours everywhere,” he says, warning that separate playbooks are “how you end up disclosing in Dusseldorf before your own board in Boston has heard about it.”
How much to say in early breach updates
Both men are wary of premature reassurance. Pontiroli notes that restoring a service does not prove an attacker has been removed, and that an extortion group’s claims should not quietly become the company’s own account of events. Maude points to the familiar line that there is “no evidence of misuse”, which too often means there are no logs. “Customers do not want to hear that you take security seriously; they want to know whether to change a password, freeze a card, or ignore the next email claiming to be from you.” His advice is to disclose in versions, setting out what is known, what is not and when the next update will come. “Nobody penalises a company for a second update that says ‘the impact is smaller than we feared’.”
Privileged access is now a board issue
Whether stricter disclosure rules make companies safer or simply more exposed to lawsuits, Maude’s answer is both, and he considers that useful. When a 72-hour deadline reveals that a company cannot say which accounts were touched or what was taken, visibility stops being an abstract debate and becomes a budget line. The rules have pushed one question into every boardroom, he says: “Do we actually know who and what has privileged access in this organisation? Most do not like the answer, which is exactly why the question matters.”
With non-human identities outnumbering staff by more than a hundred to one, the “what” in that question is now carrying most of the weight, and it may be the most useful thing for enterprises to become aware of this October.
