AI disrupts cybersecurity and pushes defenders to change methods

Alex Stamos believes AI accelerates attacks and weakens the current defense model. For him, companies will need to test their systems faster, cheaper, and with more automation.

AI is reshuffling the cybersecurity deck, to the point of rendering obsolete several habits established over the past twenty years, according to Alex Stamos, chief security officer at Cognition, in an interview published by The Verge on November 7, 2026. He describes a wave of attacks that are faster, better automated, and cheaper to launch, while defense remains costly and too slow. In this context, companies will no longer really have the luxury of waiting for the next miraculous patch.

A cyber defense too slow against attacks that automate themselves

Stamos' diagnosis is severe: the cybersecurity industry is, in his view, living through a “dark age” during which attackers gain capabilities faster than defenses can adapt. He cites a series of recent incidents, including breaches at AI-related companies, the Binance hack, and the Cursor incident, to illustrate a chain of cascading attacks. The key point is not just the number of breaches, but the speed at which they follow one another and spread.

This acceleration stems from a technical shift that is simple to describe but brutal in its effects: open-weight models can now automate a large part of the “kill chain,” that is, the steps leading from a vulnerability to an effective intrusion. Where attackers previously had to prepare their tools extensively, they can now detect a vulnerability, write the exploit, and act almost in real time. Alex Stamos sums this up as a reversal of the classic model, with systems that “go wild” on targets. The phrase is colorful, but the message is clear.

Why costs change everything in the digital war

The issue is not only technical. It is economic. Stamos emphasizes a dangerous imbalance: the cost of attack tends toward zero, or nearly so, while the cost of defense remains high. He notably cites security investigations billed at 900 dollars on certain cyber model platforms, an amount difficult for defenders to absorb when attackers work with reused hardware or nearly free resources.

In his reasoning, the former industry balance relied on a scarcity of skills. A few specialists really knew how to build offensive tools, and companies ended up structuring an industry around security, bug bounties, and responsible disclosure. This framework is cracking. Now, more profiles can access capabilities once reserved for a small elite. This does not mean everyone becomes an expert, but enough people become dangerous for the risk surface to explode. And security managers did not sign up to run a marathon in sandals.

The real challenge: securing operations, not just code

An important part of the problem, according to Stamos, remains poorly addressed: operational security. Detecting a flaw in the code is useful, but it is not enough if a company does not know how to read its logs, launch an investigation, or decide quickly without blocking all activity. He believes the first layers of the Security Operations Center must now be automated, with systems simpler than large language models to process massive volumes of logs, then more sophisticated models for analysis and response.

He also describes a shift already underway in the companies he supports: scanning repositories, discovering thousands of bugs, generating tests, then automatically merging patches. In his view, a complete overhaul sometimes becomes less costly than hunting flaws one by one. This logic also applies to historical components like the Linux kernel, OpenSSL, or glibc, which he considers too exposed to be simply “patched” indefinitely. His idea goes far: a modern fork of the kernel, limited to server hardware, rewritten in memory-safe and type-safe language. Ambitious, yes. But in the world of eternal patches, utopia has already lost some points.

Companies will have to red-team themselves

Stamos states that organizations will have to actively test their own defenses within the next 90 days, either with AI red teaming tools or with specialized providers. Otherwise, they will be tested by attackers. He believes most large groups do not yet have a risk management framework that allows this type of large-scale exercise, even as threats execute faster than internal validation circuits.

He is more cautious about the idea of a “swarm” of autonomous defensive agents launched without safeguards. Instead, he advocates systems bounded by deterministic limits: restricted access, capped number of actions, escalation to a human beyond a threshold. It is a cautious approach, but consistent with his initial observation: in cybersecurity, total autonomy is not a strategy, it is a risky bet.

The mirror of fake content and upcoming fraud

The interview also drifts toward AI-generated content, which poses a similar problem: authenticity verification. Stamos considers it more realistic to prove that content is real than to demonstrate that content is fake. For photo and video, he sees an advantage in cryptographic provenance chains, provided capture, editing, and publishing are controlled within the same environment. For text, however, the task seems much more uncertain, especially for short formats.

The issue goes beyond social networks. Phishing, voice scams, family impersonations by video: everything that relies on the appearance of reality becomes harder to distinguish. This is where AI puts the most concrete pressure on daily uses. Companies, meanwhile, will have to deal with two fronts at once: faster technical attacks and more credible fraud. In a world where parents are already asked to call before believing a video of “kidnapped grandchildren,” the level of vigilance has clearly changed.

Key points

  • Alex Stamos talks about a “dark age” for cybersecurity.
  • On November 7, 2026, The Verge publishes his interview.
  • He cites 900 dollars for an investigation on certain models.
  • AI red teaming would become necessary within 90 days.
  • Binance, Cursor, and Hugging Face serve as recent examples.
  • Cognition uses deterministic limits in its SOC, according to Stamos.

By the numbers

  • 90 days — horizon announced by Alex Stamos to red-team oneself.
  • 900 dollars — cost cited for an investigation on a cyber model, according to Stamos.
  • 5,000 companies — potential spread mentioned after the Cursor incident.
  • 700 to 1,400 agents — estimated size of the OpenAI “swarm” mentioned in the interview.
  • November 7, 2026 — date of publication of the interview by The Verge.

To read