An intrusion that once took a human attacker the better part of two working weeks can now be finished before the morning coffee goes cold. Nicole Quinn, vice president at Palo Alto Networks, told RNZ that where a hacker might once need nine days to infiltrate a system, an autonomous AI agent can do it in as little as 15 minutes. A phone vulnerability that criminals took seven to nine days to spot now takes about 22 minutes.
That changes who owns the problem. When the attack window is measured in minutes, cyber security stops being an IT maintenance line and becomes a business continuity risk that sits with the board.
Same old tricks, running at machine speed
Quinn is clear that AI isn’t inventing new attack methods. It is a “significant step change” in speed and scale, because AI “doesn’t get tired” and can chain together a string of low-severity weaknesses into one serious exploit, work that used to need human ingenuity. She also admitted regulators and industry are “still flying and building the plane at the same time”.
The vendor caveat is obvious. Palo Alto sells the defence to the threat it describes. But the independent numbers point the same way. McKinsey found mean time to exploitation for critical vulnerabilities fell from about 23 days in 2025 to an estimated 20 hours in 2026. Its diagnosis is the one every director should hear: attackers move at machine speed while organisations decide at committee speed, and that gap is the real vulnerability.
Officials are already saying the same thing
This is not a sales pitch the government is ignoring. On 20 November the NCSC warned that attacks which once took days or weeks now happen at machine speed, and that businesses need to plan for AI disruption now.
The NCSC’s annual figures look reassuring on the surface. Recorded incidents fell to 4,673 in 2025/26 from 5,995, and direct losses dropped to $23.8 million from $26.9 million. But incidents escalated for specialist support because of potential national significance rose to 369 from 331. Fewer attacks, nastier tail. The same report warns that AI agents are becoming capable of chaining vulnerabilities into complex exploits, and that by early 2027 criminals may have capabilities currently confined to frontier models.
Financial regulators have moved first. In October the Reserve Bank and FMA told regulated entities to assume AI-enabled incidents will occur, run scenario tests, set an explicit AI risk appetite and fix legacy systems. FMA executive director Louise Unger put it bluntly: “Boards and senior management need to know where AI is being used in their business” and who is accountable for it. Banks and insurers got the memo in writing. Everyone else should read it as a preview.
Half the plans have never been tested
Here is the uncomfortable bit. Kordia’s research, published in March 2026, found 44% of large New Zealand businesses had suffered a cyber attack or incident in the previous 12 months. Of those, 17% faced serious disruption, 61% had personal information accessed or stolen, and about 17% of incidents involved AI misuse.
Yet writing for the Institute of Directors, Patrick Sharp, general manager of Aura Information Security, notes that only about half of businesses had tested their incident response plans in the past year. He also points out the time to reverse-engineer a patch has collapsed to an average of five days, making monthly patching cycles obsolete. A plan that lives in a drawer is not a plan when the attacker is finished in a quarter of an hour.
Agents with keys to the building
The newer risk is inside the firewall. Palo Alto’s Riccardo Galbiati told BusinessDesk that AI agents inheriting human logins can break traditional access controls because “they try every path, every possible option” when given agency. Quinn wants every agent given a trackable identity rather than left as anonymous code.
Australia has already had its warning shot. University of Auckland academics Dr Dulani Jayasuriya and Professor Alex Sims examined how an autonomous agent gained unauthorised access to the Medicare statistics portal, and argue New Zealand needs its own expertise to test agentic systems rather than relying on overseas developers’ assurances. That applies equally to vendor assurances in your next procurement round.
What boards should do before Christmas
None of this requires new legislation. It requires directors to treat speed as the risk. That means a written AI risk appetite, an inventory of every agent and connector touching company systems, identity controls on those agents, patching measured in days rather than months, and an incident response plan rehearsed against a 15-minute scenario, not a nine-day one.
The regulators have told finance boards to get there. The NCSC has told everyone else. With criminals expected to gain frontier-grade tools by early 2027, the businesses still scheduling their cyber review for next year’s audit committee are betting they won’t be the 15-minute case study.
Sources
- RNZ: Cyber attacks can happen within minutes thanks to AI (2026-11-21)
- McKinsey: How Frontier AI is compressing cybersecurity response times (2026-09-11)
- RNZ: Financial regulators urge stronger cyber security amid AI-enabled cyber attacks (2026-10-16)
- Kordia: New Zealand Business Cyber Security Report 2026 (2026-03)
- Institute of Directors: AI is shrinking the cyber defence window (2026-06-25)
- BusinessDesk: Palo Alto Networks says AI agents can exploit inherited logins (2026-11-11)
- University of Auckland: Australia’s Medicare breach a warning for Aotearoa (2026-10-07)
Join the discussion
Add useful context, ask a good question, or challenge an idea — keep it specific and respectful.