AI hacking threatens the internet, but recent outages prove nothing yet

AI systems that find and exploit software flaws could outpace the people who patch them. That risk is real, but the run of outages in the past week is.

AI systems that find and exploit software flaws could outpace the people who patch them. That risk is real, but the run of outages in the past week is not evidence of it, and the connection between the two is still unproven.

Key takeaways

  • The central worry is speed: advanced AI models may find and exploit software vulnerabilities faster than organisations can patch them.
  • The Guardian reports that recent outages at Microsoft, United Airlines and JPMorgan Chase came with official statements that there was no evidence of a cyber-attack.
  • These outages show how fragile and interconnected digital infrastructure has become, which matters whatever the cause of any single failure.
  • The same AI capabilities that help attackers can also help defenders, and nobody yet knows which side will gain more.

The argument: speed is the new vulnerability

A Guardian opinion column makes the case. Frontier AI models, meaning the most capable general-purpose systems from the leading developers, may soon find and exploit flaws in software faster than the world can fix them. The column also argues that the risks reach beyond computer security.

The argument does not depend on any one dramatic attack. It rests on an imbalance between attack and repair that already exists. Finding a flaw and fixing it are separate jobs, done by different people at different speeds. Once someone discovers a vulnerability, a working exploit can be written relatively quickly. A fix has to be written, tested, distributed and installed across every affected system. In large organisations that can mean old equipment, systems that cannot easily be taken offline, and software whose original authors are long gone.

That gap is not new. Security teams have long described their work as a race between disclosure and deployment of patches. The new claim is that AI could make the attacking side of the race much faster while the defending side stays slow. Patching is limited by organisational processes, testing requirements and physical infrastructure, and better algorithms do not remove those limits.

If that is right, the problem is structural. It would not be solved by stopping a particular group of attackers. It would mean that the cost of finding exploitable weaknesses falls sharply while the cost of fixing them hardly changes.

Evidence: discovery scales with computing power, repair scales with people

The first piece of supporting evidence is the different ways the two activities grow.

Searching for vulnerabilities is largely a matter of reading code, reasoning about how it behaves and testing unusual inputs. Automated tools have helped with parts of this for years. Language models add the ability to read and interpret large codebases in a way that looks more like a human analyst’s work. A system that can do this can, in principle, be copied and run in parallel on as many targets as its operators can pay for. Capacity grows with computing resources.

Repair does not grow that way. A patch for widely used software usually has to be written by a small group of maintainers, some of them volunteers working on open-source projects. It then passes through vendors, system administrators and change-control procedures before it reaches the machines that need it. Hospitals, utilities, airlines and banks often run systems where an untested update could cause more disruption than the flaw it fixes. Those organisations patch cautiously, for good reasons.

So even if defenders use the same AI tools to find flaws first, finding them is only the first step. A list of a thousand newly discovered vulnerabilities is useful only if someone has the capacity to fix them. If AI speeds up discovery for both sides while the human, organisational work of remediation stays the same, the backlog of known but unfixed weaknesses could grow. That backlog is what attackers use.

This publication cannot verify precise measures of how much faster current models make vulnerability discovery. The Guardian column does not give such figures in the material available here, and none are offered in this article.

Evidence: recent outages expose a brittle system

The second piece of evidence is not about AI. It shows how much damage a failure in core digital services can do.

The Guardian reports that in the past week outages hit core services at Microsoft, United Airlines and JPMorgan Chase. On Friday, according to the same report, all domestic departures in the United States were grounded, the second time this has happened since the attacks of 11 September 2001. The Guardian also cites CNN reporting that the National Security Agency’s highly classified internal communications network went down on Tuesday.

The Guardian says that in each case officials or the companies involved stated there was no evidence of a cyber-attack. That point matters, and it is easy to lose when such events are grouped together. On the information available, none of these failures can be attributed to AI or to any hostile actor. Their causes, as far as this publication can establish, have not been made fully public.

Still, they are relevant to the wider argument. They show that a small number of software platforms and service providers sit underneath large parts of daily life, including air travel, banking and government communications. When one of them fails, for whatever reason, the effects spread quickly and widely. A system where accidental faults can ground a nation’s flights is one where a deliberate, well-targeted exploit could do at least as much harm.

The argument, then, is not that AI caused last week’s disruption. It is that the disruption shows the stakes. Infrastructure that fails this visibly under ordinary stress would be badly exposed if the supply of exploitable flaws rose sharply.

Evidence: the same capabilities reach beyond software

The third strand is the column’s claim that the risks extend beyond the cyber realm. The material available here does not set out that wider case in detail, so it can only be described in general terms.

The underlying reasoning is familiar from debates about advanced AI. A system good enough to reason through complex code may also be good at other technical problems where a capability can help or harm, such as chemistry, biology or the control systems that run physical infrastructure. Software flaws are an early test because the domain is entirely digital. Results can be checked automatically, and a successful exploit gives a clear signal. That makes cyber security a likely place for dangerous capabilities to appear first.

There is also a direct link between the digital and the physical. Power grids, water treatment, transport signalling and medical devices all depend on software. An exploit in those systems is not limited to stolen data or disrupted services. It can affect physical equipment and public safety. Seen this way, “breaking the internet” is shorthand for something broader: weakening the computer-controlled systems that modern societies rely on.

The case against: defenders gain too, and outages are mostly mundane

The strongest counter-argument starts with the fact that AI tools are not available only to attackers. The organisations best placed to use advanced models at scale include large technology companies, security vendors and government agencies, which have strong reasons to use them defensively. AI-assisted code review could catch flaws before software ships, which removes the need for a patch at all. Automated tools could also help write and test fixes, which would shorten the slow part of the cycle that the pessimistic argument relies on.

On this view, the likeliest outcome is not a lasting advantage for attackers but a period of upheaval. Many old flaws are found and fixed, and new software then becomes harder to break because it was checked with the same tools from the start. Earlier waves of automated security testing were also expected to favour attackers, and they ended up as standard defensive practice.

A second line of argument concerns the outages themselves. Large-scale IT failures have many ordinary causes: faulty updates, configuration errors, hardware faults, capacity problems and dependencies between suppliers. In the cases the Guardian describes, the organisations involved said they found no evidence of an attack. Treating a cluster of such failures as a sign of AI-driven danger risks seeing a pattern that is not there, and could push policy towards a dramatic but unproven threat and away from more routine weaknesses that are known.

Third, the most capable models are made by a fairly small number of developers who can restrict access, monitor use and refuse to help with obvious attack requests. Those safeguards are imperfect and can be bypassed, but they add friction that the bleakest scenarios may underrate. Real attacks also need more than a vulnerability. They need access, persistence, a way to avoid detection and a motive. AI may speed up one stage of that chain without changing the others much.

Taken together, this is a serious case. It does not show that the risk is imaginary. It does suggest that the result depends on choices about deployment, investment and regulation, and is not fixed by the technology alone.

Evidence that would change the conclusion

Several kinds of evidence would make the pessimistic view more convincing or less.

Investigations into recent major outages that confirmed an AI-assisted attack would be significant, as would documented cases of AI being used in serious attacks on critical infrastructure. At present, the Guardian reports official statements that no cyber-attack was found in the cases it lists.

Systematic data on the time from vulnerability disclosure to exploitation, and from disclosure to widespread patching, would show whether the gap is actually widening. If exploitation times shorten sharply while patching times stay flat, the argument is strengthened. If AI-assisted tools are shown to shorten patching times as well, it is weakened.

Independent evaluations of what frontier models can do in realistic security tests, published openly and repeated over time, would show how fast offensive capability is growing. So would evidence on whether developers’ safeguards are holding up in practice.

Finally, evidence of how organisations are responding matters. That includes whether maintainers of critical open-source software get more resources, whether operators of essential services modernise their legacy systems, and whether regulators set requirements for secure development. Each of these would narrow the gap that the warning depends on.

Until such evidence is available, the most defensible position is a middle one. The risk of AI-accelerated exploitation is plausible and serious enough to plan for now, but the recent outages are best read as evidence of fragile infrastructure, not of an attack already under way.

Sources and further reading

  • The Guardian: opinion column on AI and internet security, including its account of recent outages and official statements about their causes.
  • CNN: reporting, cited by the Guardian, on an outage affecting a classified internal communications network at the National Security Agency.
  • National cyber security agencies: general published guidance on vulnerability disclosure, patch management and protecting critical infrastructure.
  • AI developers’ published safety evaluations: general material on how frontier models are tested for cyber security and other dual-use capabilities.

Surfaced from the rss:guardian_tech signal “AI cyber risk commentary”. AI-assisted draft, editorially reviewed.

Visited 1 times, 1 visit(s) today
share this recipe:
Facebook
X
WhatsApp
Telegram
Email
Reddit