Signals/20 Nov 2026
AI speeds up software flaw discovery and cyberattacks
Nature reports that researchers say AI now finds software flaws faster than humans, shortening the gap between disclosure and attempted exploitation to a few days and making defenders' work harder.
In plain words
Nature reports that artificial intelligence now lets people find weaknesses in software much faster than before. Attackers and defenders both use it. Defenders have less time to fix a problem once it is made public. Most break-ins still start with simple human mistakes, such as default passwords. Changing default passwords and installing updates promptly remain the most useful steps for ordinary users and small organisations.
What happened
Nature's report of 19 November 2026 links several recent events. It cites the disruption of systems run by Microsoft, US air traffic control, the bank JPMorgan Chase and the US National Security Agency. It also cites researchers' reports of AI-assisted discovery of security flaws in Java and PHP software, and hackers using AI tools to target supermarket fridges and freezers in Europe. Nature calls the incidents "probably unrelated".
The researchers quoted say AI is changing how software vulnerabilities are found. Steven Murdoch of University College London says "I think vulnerability discovery without AI is basically dead"1. Agentic systems use automated tools that can search millions of lines of code. Nature says Google has reported that attackers have already used agentic AI to examine software updates, identify the security problems they fix, and find ways to exploit them.
In the Java case, Nature says AI identified circumstances that could let an attacker corrupt computer memory. Murdoch says the significance has been exaggerated and that it is not among the top ten vulnerabilities this year. He also says AI has shortened the window between public disclosure and attempted exploitation from several weeks to a few days.
What is confirmed and what is not
What the report establishes, through named experts and attribution to Google:
- Researchers say AI is widely used for vulnerability discovery, by attackers and defenders.
- Google reports attackers using agentic AI to study software updates and work out exploits.
- Murdoch says the disclosure-to-exploitation window has fallen from weeks to days.
- Daniel Gardham of the University of Surrey says AI tends to find the same kinds of weakness as human hackers, but lets people with limited expertise break in more easily.
What is only claimed or uncertain:
- Murdoch says the outages at Microsoft and elsewhere are likely technical malfunctions, not a large-scale cyberattack. The report gives no investigation findings.
- The report gives no measurements behind the "weeks to days" figure.
- It does not say how many attacks use AI, or how much of the Java and PHP discovery work was done by attackers rather than defenders.
- Murdoch says the most accomplished human researchers remain better at finding entirely new types of weakness. The report does not test this.
- 1Security teams and software maintainers racing to patch flaws9
- 2Operators of large online services and critical systems7
- 3Small organisations with default-password devices6
- 4People targeted by personalised phishing messages5
- 5Vulnerability researchers not using AI tools4
Analysis
This is a single report from one outlet, built mainly on interviews with two academics and one reference to Google. Each claim below stays within what it supports, and the primary documents are named at the end.
Defenders have less time to patch, and rushed patches can cause outages
The most direct consequence in the report is the squeeze on the time between a flaw becoming public and someone trying to use it. The mechanism, as Nature describes it: researchers normally tell a developer privately, the developer prepares a fix, and details are then published. If attackers can feed the fix itself into an AI agent and have it work out what was wrong and how to exploit it, then publication of a patch becomes a starting signal for attackers. Murdoch's figure of a few days, down from weeks, would leave organisations little room to test updates.
The report adds a second pressure. A rushed patch can interfere with other programs and cause the kind of outage that organisations are trying to avoid. Toswo Research's reading is that this creates a real trade-off for operators of large systems: patch fast and risk breaking something, or test and stay exposed. How sure is this? The direction is plausible and follows from the sources, but the size of the effect is unknown, because the figure comes from one expert's statement without data.
Technology and industry
Murdoch says vulnerability research without AI is effectively finished, and that researchers who do not use it risk falling behind. That affects security teams, which can now examine much larger amounts of software, and it affects developers, who may receive more reports. Nature also says researchers are building AI agents that detect and repair vulnerabilities automatically. The open question, stated in the report, is how thoroughly AI can check software and whether it could introduce new weaknesses. No result is given, so that risk is unquantified.
Security and who can attack
Gardham's point concerns scale and skill. He says AI finds the same kinds of weaknesses as human hackers, but a person with limited expertise can start an agent, return in 20 minutes and have dozens of results. If that holds, the number of people able to probe systems grows even if the type of attack does not change. Murdoch's counterpoint is that AI is best at familiar flaws and the best humans still lead on new types. Taken together, the report suggests more routine attacks and no established change in the most advanced ones. A sceptical expert would ask for data on how many of the flaws found were already known categories.
The human element remains central. Nature says most breaches still involve human failure. In the European supermarket case, a common problem was default passwords such as 'admin' or '12345'. AI-written messages can also draw on public information to personalise phishing and make it look authentic. Here the link between AI and harm is indirect: AI raises the quality of the lure, but the breach still depends on someone clicking or on a weak password.
Daily life and critical infrastructure
The report opens with outages at Microsoft, air traffic control, JPMorgan Chase and the NSA, but it does not claim AI caused them. Murdoch says they are likely technical malfunctions. Readers should therefore not treat those outages as proof of AI-driven attacks. They do show how much depends on a few digital systems. Toswo Research has covered one such failure in its note on the Microsoft 365 outage, and the same dependence is why preparation guidance such as the Day Zero plan exists. The supermarket fridge case is covered in a separate signal on France blaming Iran for the fridge hack; the Nature piece does not attribute that attack to any state.
The strongest reason this view could be wrong
The central claims rest on expert opinion. If measured data showed that exploitation windows had not shortened, or that most AI-found flaws were minor like the Java case Murdoch describes, then the alarm would be overstated. The opposite is also possible: Google's report may describe only the cases it could see. The report states that the recent incidents are probably unrelated, so any link between them and AI is not supported.
What would settle it
The Google report on attackers using agentic AI to examine software updates would show how common this is and what it found. Published data from vulnerability trackers on time from disclosure to first exploitation would test the weeks-to-days claim. The advisories for the Java and PHP flaws would show their severity. Official post-incident reports from Microsoft, the air traffic authorities, JPMorgan Chase and the NSA would show whether the outages were malfunctions or attacks. A Toswo Research view of how such risks are tracked is in the method note.
What can be done
For individuals and small organisations: change default passwords on any device connected to the internet, including freezers and cameras. Install updates promptly. Treat unexpected links with suspicion, including messages that look personal.
For organisations: shorten patch testing cycles where possible, and plan for the risk that a patch causes its own outage. Security teams can use AI tools to examine their own code, as Murdoch says researchers increasingly do. For software developers, Nature says researchers are working on agents that detect and repair flaws automatically, though no results are given. No single fix is described in the report.
Questions readers ask
Is AI causing the recent Microsoft and air traffic outages?
Nature does not say so. Murdoch says it is unlikely the recent outages were connected as a large-scale cyberattack and that they are likely technical malfunctions. The report gives no official cause.
How fast can attackers use a newly published software flaw?
Murdoch says the window between public disclosure and attempted exploitation has shrunk from several weeks to a few days. The report offers no data behind the figure.
Does AI help beginners hack?
Gardham says AI makes it easier for people with limited technical expertise to break into systems. He says someone can start an AI agent and return in 20 minutes to dozens of results. He adds that AI tends to find the same kinds of weakness as human hackers.
What is the simplest protection?
Nature says most breaches still involve human failure such as weak passwords or clicking unfamiliar links. Default passwords were a common problem in the European supermarket freezer case.
Quoted from
- AI is turbocharging cyberattacks — can computer scientists catch up?, Nature, 2026-11-19
Written by Toswo Research. This is not investment, legal or medical advice. To report an error, see corrections.
Toswo Signals, "AI speeds up software flaw discovery and cyberattacks", https://www.toswo.com/signals/ai-speeds-up-software-flaw-discovery-and-cyberattacks, published 2026-11-20, retrieved 2026-11-22.