Latest

Technology

AI shatters the myth that Linux can't be infected by exploits created in minutes

We analyze the myth and the reality of Linux security: robust by design, but not infallible or immune to AI-accelerated exploits

Is Linux a secure system? The myth and the reality
Is Linux a secure system? The myth and the reality

For decades, Linux has enjoyed an almost bulletproof reputation, one that made it clear that this is the system that "doesn't get infected," that runs without antivirus and silently underpins much of the Internet. A reputation with strong foundations, but also with major nuances.

Just to give one example, in 2024 the Linux kernel accumulated more than 3,000 CVEs, the highest figure in its history. In fact, in 2025 and so far in 2026, AI has made malware and exploits targeting Linux grow at a pace that no traditional patching model can keep up with.

A reality that has begun to exert pressure and leads us to the question: Under what conditions is Linux really secure?

Linux is very secure by design

Answering that question brings us to a first point: What makes Linux secure? And the initial answer is clear: the way it works and is built.

At this point, there are structural reasons that give Linux an advantage. The first is its permissions model: users work with limited privileges and any system change requires explicitly elevating to root (administrator). Thus, for example, malware that compromises a session is initially trapped within its perimeter, with no global damage.

In Windows, this is not so clear. For years, Windows had no accounts other than the administrative one, and with the introduction of accounts, the limitations were not very clear.

It was not until Windows 10 that this privilege separation became secure enough to be considered a secure barrier, except that nobody uses it. Most people who use Windows do so with administrative accounts, so it is useless. In Linux, that is unthinkable. In fact, many applications cannot run as root, they simply fail, a measure in favor of maintaining security.

Vitalik explores a new cryptography resistant to quantum computers

The advantage of free software

The second is open source. The Linux kernel currently exceeds 36 million lines and has received contributions from more than 13,500 developers since 2005. Linus's Law, which says: "with enough eyes, all bugs are shallow," means that an audit is produced that no proprietary vendor can replicate at that scale. A review that allows flaws to be exposed earlier and makes every patch a public good.

Added to this are unusual hardening layers: SELinux's mandatory access control (MAC) (created by the NSA) and AppArmor limit what each process can touch even if it is compromised; seccomp filters system calls; and installation goes through package managers with signed repositories. With auditing, encryption and updates in a couple of commands, a well-administered desktop does not need antivirus.

There is also an economic factor. More than 95% of known malware targets Windows (AV Atlas), because that is where the mass of users with exploitable data is, while Linux barely exceeds 7% desktop share (StatCounter) and is fragmented across hundreds of distributions. Attacking the Linux desktop has not been profitable, nor is it simple.

The other side of the coin: the kernel and the mountain of CVEs

But that design advantage coexists with an uncomfortable fact: Linux is the open source component with the most vulnerabilities recorded in NIST's national vulnerability database (NVD). Between 1999 and 2024, more than 20,000 CVEs were assigned to the kernel, and the pace has skyrocketed in recent years.

But these figures should be read in context. In February 2024, the kernel assumed its role as a CVE Numbering Authority (CNA) and began assigning identifiers to almost all security fixes in its stable branches, even minor ones.

The 2024 jump is more transparency than more bugs, something Greg Kroah-Hartman announced while warning of an "apparently enormous" number of CVEs. The exact figure varies depending on how it is counted. The NVD database records 3,108 CVEs in 2024 and other readings of the same data, such as Swif.ai's, reach 3,529. That is a rate of 8-9 new CVEs per day, for the kernel alone.

Increasingly serious problems

But the most worrying thing is not the volume, but the severity. Critical CVEs (CVSS 9.0+) went from 22 in 2019 to 148 in 2024, that is one every 2.5 days, and 42% of the total was classified as high severity. Subsystems such as networking and memory management generated 38% of CVEs between 2020 and 2024, dominated by memory safety flaws (use after free, out-of-bounds write). This does not contradict the soundness of the design: it is a reminder that the kernel is mostly written in C, without memory guarantees.

The comparison reinforces the nuance. In 2024, the kernel recorded 3,108 CVEs compared with 587 for the Windows kernel or 312 for Mac OS, with a similar percentage of critical and high severity (46.8% versus 54.2% and 51%). The volume gap reflects the open model, but also that Linux is not immune to serious flaws, and in fact the level is very similar to what we see in Windows and Mac OS.

Microsoft's 18 rules for human-AI interaction

The myths worth debunking

And this now brings us to certain myths worth debunking:

  • The first: "Linux is not prone to viruses". Classic viruses are rare on Linux, it is true, but worms and ransomware attack any system with value, as security expert Michael Boelen at Linux Audit, points out. The idea that "nobody is interested in Linux" became obsolete when Linux became the system of servers, the cloud and IoT.
  • The second: "I don't need a firewall". A well-configured distribution exposes few services, but even so, a system without network filtering is a bad citizen of that network. After all, the firewall remains a basic defense against worms and network attacks, inbound and outbound. Curiously, this is the most common mistake in the Linux world, since most Linux systems come without a properly configured firewall.
  • The third: "Open source is secure by definition". The availability of the code does not guarantee its review. Linus's Law works when there are plenty of eyes, and not all projects have them. An example of this is XZ Utils, where a malicious actor was able to inject a payload to infiltrate Linux systems worldwide, all because there were not enough developers and eyes to review the code.
  • The fourth: "Packages from official repositories are secure". This is the other side of the error. An official repository contains thousands of packages, and a distribution maintaining it means it patches known flaws, not that they do not exist, or that they are not introduced by the maintainers themselves. At this point, Arch Linux is the distro with the most historical problems detected and that are still occurring.

A serious problem

The most dramatic demonstration was precisely the XZ Utils flaw (CVE-2024-3094), in March 2024. An actor under the alias "Jia Tan" spent more than two years contributing legitimate patches to the compression library, gained the community's trust and was named maintainer.

In versions 5.6.0 and 5.6.1 of this software, a backdoor was introduced that altered OpenSSH authentication and allowed remote execution before authentication; CISA rated it with the maximum CVSS 10.0. It was discovered by chance by Microsoft engineer Andres Freund, when he noticed anomalous CPU consumption in SSH logins. Stable distributions (RHEL, Ubuntu LTS, Debian stable) were saved by days. Because had the whole process been completed, millions of systems would have been within reach of the attack.

Another illustrative case is Looney Tunables (CVE-2023-4911), revealed by Qualys in October 2023: a buffer overflow in the glibc dynamic loader that granted root privileges on default installations of Fedora, Ubuntu and Debian. CISA added it to its catalog of exploited vulnerabilities in November 2023 and it remained an active target throughout 2025.

The target is no longer the desktop

To all this can be added another reality: according to W3Techs, 62% of websites with an identifiable operating system run on Linux as of August 2026, 96% of the top million web servers and the 500 fastest supercomputers use Linux, and 91.6% of Google Cloud virtual machines and 83.5% of AWS ones run on it. Android, with more than 3 billion devices, is also Linux.

When critical infrastructure concentrates on a single system, that system becomes the target, and the data from 2025 and 2026 confirms it starkly:

  • Brute force by volume. The Elastic global threat report (2025) found that 89% of endpoint behaviors observed on Linux systems were brute-force attempts, especially against exposed SSH: attackers sweep Linux with credentials every minute.
  • Ransomware specific to Linux. LockBit, Play, Akira and the newer Kraken family distribute Linux encryptors dedicated to VMware ESXi hosts: a single compromised hypervisor encrypts dozens of virtual machines, and that is why this ransomware accounts for a disproportionate share of high losses. CISA added ESXi's arbitrary write flaw CVE-2025-22225 to its catalog of exploited vulnerabilities in 2025.
  • IoT botnets. Cloudflare absorbed a DDoS attack with a peak of 31.4 Tbps attributed to the Aisuru-Kimwolf variant of Mirai, which is estimated to have compromised between one and four million Linux devices.
  • Cloud and containers. Cloud intrusions grew 26% year over year and attacks 37% (CrowdStrike), with abuse of valid credentials as the main method. In a Sysdig case, attackers launched more than 500 cryptomining containers every 20 seconds, and a public cloud breach costs an average of more than 5 million dollars.

AI has accelerated the attack against Linux

And here the debate changes in nature. AI has not created a new category of vulnerabilities in Linux; it has compressed the time between disclosure and exploitation, and has made Linux-specific malware cheaper, to a point that no traditional patching schedule can assume.

The Cloud Security Alliance (CSA) documented in April 2026 the collapse of the "exploitation window." The average time between disclosure and exploitation fell from about 32 days in 2022 to about 5. And, in 2025, 32.1% of new exploits appeared on the same day as or before the public disclosure of the CVE.

All because AI systems today generate PoCs for published CVEs in 10-15 minutes at a cost of one dollar per attempt; multi-agent frameworks such as CVE-Genie reproduced 51% of the 2024-2025 CVEs, with verifiable exploits, at an average cost of 2.77 dollars. Anthropic's 2026 research showed models that built Firefox exploits in less than an hour from the patch's publication and complete Windows kernel privilege escalation chains in hours.

DFlash 2: the technology that lets AI respond three times faster

Multivector and hybrid attacks

In January 2026, Check Point Research presented VoidLink, the first Linux malware framework of advanced complexity developed almost entirely by AI: some 88,000 lines of code in less than a week, by a single operator using the AI-assisted IDE, with more than 30 modules, hybrid rootkits (LKM and eBPF) and cloud-native persistence.

Elastic Security Labs corroborated in March 2026, on the leaked source code, iteration cycles typical of conversations with language models, and warned that the barrier to developing kernel rootkits has dropped drastically. In August 2026, The Hacker News reported on 14 trojanized npm packages that distributed the RedC2 4.0 backdoor for Linux with AI-assisted C2 capabilities.

Autonomous attacks have reached the same level. Unit 42 documented in July 2026 a campaign by a Chinese-speaking actor who used DeepSeek through the Hermes Agent framework to enumerate targets, scan GitHub for trending exploits and select vulnerabilities by severity: more than 460 targets, executing hundreds of hours of manual work in minutes.

In parallel, academic research described an autonomous worm that runs language models on the stolen computing power of the systems it infects, replicates and exploits vulnerabilities published after the model's training cutoff, turning public advisories into exploits without human intervention.

The verdict

Is Linux a secure system? Yes, more than most, by design and by history. Is it infallible? No, and the evidence of the last three years demonstrates it unambiguously. The myth that "Linux is not attacked" died when Linux became the operating system of the world's data. The reality is far more nuanced: Linux remains a solid platform, but its security today depends on what its administrators do, not on what the logo promises.

With AI increasingly compressing the window between disclosure and exploitation, the difference between a compromised system and a safe one will not be the system, but the discipline with which it is operated.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading