Skip to content

AI in Cybersec

Machine Speed is a lie: stop trying to fight AI with AI

A masterpiece, by Marcus Hutchins: "In 2017, I stopped WannaCry, an autonomous ransomware worm that infected millions of systems." (September 2026)

  • We’re not on the cusp of a cyber-apocalypse, but a continuation of a multi-decade long trend where attack techniques continue to evolve, improve, and accelerate year-over-year.
  • Generative AI might be faster than humans at many tasks, but it’s also one of the slowest forms of automation...In the time it takes you to get ChatGPT to respond to a simple “hello”, WannaCry could have infected hundreds of systems. Most threat actors also aren’t interested in operating on the “cutting edge” of technology...They follow the path of least resistance. What’s good enough, doesn’t require too much effort, isn’t expensive, or hard to maintain.
  • While initial access is often the result of automated hacking campaigns, the privilege escalation and lateral movement portion is primarily conducted by human operators...It is my belief that most threat actors have neglected to automate these kinds of hands-on-keyboard attacks because, simply, they’re winning consistently already. They’re ransoming networks at a high enough success rate that it doesn’t make sense to invest the time and money into building out automations.
  • In spite of the network lacking any real defenses, Mythos failed 70% of the time, with a budget of 100 million tokens per attempt. In terms of real cost, that’s $2,500–$12,500 per attempt (with the exact amount depending on the ratio of input to output tokens, since output tokens are more expensive), which works out around $8,000–$42,000 per successful compromise...Mythos Preview was widely considered to be the pinnacle of offensive AI models. So its failure rate on a deliberately vulnerable undefended network should have thrown a lot of cold water on the idea that attackers could easily just pick up an LLM and supercharge their intrusions. Still, the narratives persisted.
  • Breakout times (the average amount of time it takes an attacker to go from an initial foothold, to another endpoint, or higher privileged access) have been falling at a steady rate for years, predating AI...The average breakout time was 29 minutes in 2025, and has likely fallen since then. Assuming an alert is raised the second an attacker enters your network, response times of any longer than 29 minutes fail against the average attacker.
  • Significant time often elapses between the SOC declaring a critical incident and remediation being completed. 29-minute average breakout times don’t mean the SOC has 29 minutes to triage the alert and raise an incident. It means the organization has less than 29 minutes to go from first alert to containment.
  • Defensive AIs need to consume orders of magnitude more signals when compared with attacking ones. The attacker’s AI only cares about getting to the goal as quickly as possible. It also enjoys mostly binary results. Either an attack stage succeeded, or it didn’t. On the other hand, the defender’s AI has to triage alerts, cross-reference data, figure out where the attacker is within the network, what they’ve obtained, which mitigations or remediation to deploy, and if containment was successful. They also have to care about false positives, service availability, and laws/compliance.
  • Many categories of security alert, like those related to credential dumping, have low false positive rates. For most of them, they can and should be configured to trigger an immediate endpoint quarantine and credential lockout...This doesn’t require Generative AI or fancy ML models, it’s a default feature of pretty much every EDR. There’s plenty of alert types which should invoke a “shoot first, ask questions later” response...Historically, organizations have chosen to prioritize business function and availability over security. As attacks and attackers get faster, this becomes increasingly untenable. If a threat actor can move from initial access to a full network-wide ransom in minutes, one has to ask what any single employee is doing in those few minutes that’s going to offset the cost?
  • AI SOC is sold as a silver bullet for “machine speed” attacks, but it’s a band-aid at best. The first issue is that most organization are using SOC as a substitute for effective proactive security...There already exists many tried and tested means by which to prevent attacks in the first place...Proactive solutions are going to be infinitely faster than the fastest EDR alert, and infinity * infinity faster than any Generative AI solution.
  • Features like Credential Guard protect the credentials and other secrets previously stored by lsass.exe, by moving them into a secure virtual enclave...Not only do you still get the benefits of being able to have the EDR alert on the attempt and automatically isolate the system, but the attacker doesn’t obtain anything useful in the process. No amount of AI is going to outrun credentials not being accessible in the first place...it’s also rarely required for most endpoints within a network to be directly talking to each other in the first place.
  • Attacks continue to get faster and more frequent, regardless of AI. The threat of “machine speed” autonomous attacks suddenly rampaging through the internet is nothing but marketing hype. It’s entirely contingent on people not understanding the current or historical threat landscape.

Beyond the Hype: AI, Ransomware and Business Models

Will Lyne and Jamie MacColl, Royal United Services Institute, August 27, 2026 (link).

We are frequently told by cyber security marketers that we are in the midst of an AI-powered revolution for cybercrime. The short version of these claims is that AI will significantly lower the barrier of entry for would-be cybercriminals, allow them to create autonomous variants of ransomware that will allow them to overwhelm defenders at a scale previously thought impossible.

However, such assessments start from a false premise about what drives the behaviour of cybercriminals. The history of modern cybercrime demonstrates two things. First, that cybercriminal behaviour has been driven more by innovation in business models than technical capabilities. Second, that cybercriminals tend to innovate when they have to, not simply because a new kind of technology becomes available.

VulnCheck State of Exploitation 1H-2026

Report, July 2026.

  • Anthropic reported more than 23,000 findings through Project Glasswing, but only 126 have resulted in published CVEs, and just one has been confirmed as exploited in the wild.
  • Of 1,061 vulnerabilities attributed to AI-assisted discovery, only 14, or 1.3%, have been confirmed as exploited in the wild, roughly matching the overall exploitation rate of all vulnerabilities in the first six months of the year.
  • AI products are emerging as a new attack surface, with known exploitation affecting model-building tools, workload-scaling platforms, AI gateways, agents and workflow automation.
  • The KEV-to-CVE ratio has declined over the past two years, peaking at 2.7% in the second half of 2023 and falling to 1.4% in the first half of 2026. While the first half of 2026 saw a 10% increase in KEVs compared to the prior six months, CVE volume grew at a much faster rate of 45%, resulting in a significant drop in the KEV-to-CVE ratio. Of course, exploitation often occurs months or even years after a vulnerability is disclosed, so it's still too early to determine whether exploitation volumes will eventually follow the same growth trend as CVE issuance or level off at current rates.