This advance shows that AI is not here to replace rigorous systems engineering or advanced administration, but to enhance it, solving problems that remained hidden in the deepest corners of the source code and in system scaling policies.
Auditing and generating code like never before
The impact of AI on auditing mature codebases is perhaps its most forceful demonstration of value, evidenced by the discovery of thousands of vulnerabilities in systems that were considered impregnable strongholds. Impressive cases such as the disclosure of more than 400 CVEs in the Linux kernel and the discovery of more than 300 bugs in OpenBSD demonstrate astonishing analytical capability.
This new generation of tools goes beyond simple scanning for known patterns, applying deep reasoning capabilities that expose structural flaws that have evaded the human eye for decades.
In fact, AI models such as Anthropic's Mythos managed to expose a 27-year-old vulnerability in OpenBSD, which was capable of remotely locking up machines simply by initiating a connection. The sophistication of these models lies in their ability to "chain vulnerabilities," an autonomous process in which AI connects seemingly minor software flaws to forge a cohesive and lethal attack, managing to escalate from standard user access to full system control.
AI + Blockchain 2026: Wallets, agents and DePIN networks create the infrastructure of the future
Unraveling binaries and source alike
Moreover, these AI tools not only analyze the text of the source code, but also possess the disruptive ability to examine compiled binary code directly. This skill is essential for auditing legacy systems or modules where the original code has been lost or is inaccessible, drastically expanding the evaluation surface.
Added to this is the intense debate in the open source software community over the recent proposal to include an ext4 driver in OpenBSD developed primarily by AI. A step that challenges traditional notions of code authorship and maintainability, but underlines the undeniable irruption of AI at the core of operating systems.
The copyright problem
The controversy revolves around substantial legal risks, particularly the fear of "GPL contamination," where AI-generated code could accidentally derive from protected sources, jeopardizing the integrity and legal distribution of the entire project. Simultaneously, the problem of technical sustainability arises; developers warn that the lack of a human author who deeply understands the underlying logic makes the task of auditing, debugging and adapting the code to future architectural changes enormously difficult.
This scenario forces the open source ecosystem to redefine its review standards, demonstrating that while AI is capable of generating and analyzing critical components at an unusual speed, the industry must still build the legal and technical frameworks to absorb this revolution safely. Recognizing this overwhelming asymmetry, corporations such as Anthropic have begun to allocate millions in investment to help open source foundations strengthen their defenses, on the premise that if attackers operate at machine speed, defenders can no longer rely exclusively on human methods.
FreeBSD, a front-line witness
But perhaps one of the main systems to see all this from the front line is FreeBSD. In a systematic effort to support open source teams, researchers used AI agents to audit the system, which made it possible to discover 15 critical bugs in the FreeBSD kernel in a matter of weeks.
This striking finding included 3 remote code execution (RCE) vulnerabilities, 5 local privilege escalations (LPE), memory disclosures, denials of service and a critical escape from the bhyve hypervisor from the guest to the host.
AI not only identified these flaws, but autonomously drafted working exploits and detailed technical reports for complex vulnerabilities such as setcred (CVE-2026-45250), ptrace (CVE-2026-45253) and procdesc (CVE-2026-45251). By providing concise reports and suggested patches, artificial intelligence enabled the FreeBSD maintainers' team to go from receiving the report to implementing the fix in just a few days. In a parallel incident that illustrates the astonishing speed of these systems, an autonomous agent powered by the Claude model managed to analyze and exploit a remote execution vulnerability (CVE-2026-4747) in the RPCSEC_GSS module of the FreeBSD kernel in just four hours of compute time.
Multi-packet strategy
This module, responsible for handling Kerberos authentication for NFS servers, had a flaw that allowed a malicious client to trigger a stack buffer overflow without needing to authenticate beforehand. To turn this bug into a working weapon, the model solved six highly intricate technical problems entirely on its own.
The agent set up a test environment with the vulnerable kernel, devised a multi-packet strategy to deliver shellcode that exceeded the capacity of a single packet, and managed to cleanly terminate the hijacked kernel threads to ensure the server remained operational between attack rounds.
The war for free AI: Humanity's heritage on the brink of extinction
Making the difficult easy
In addition, the AI corrected incorrect stack offsets using De Bruijn patterns, created a new process from the kernel context to make a safe transition to user space, and cleaned up legacy debug registers that caused child processes to crash, finally opening a console with root privileges. A process that for a normal person could easily take at least 15 days of work, the AI did in a couple of hours, requiring only subsequent human verification to check that everything was correct.
This level of interaction shows that operating system hardening now requires AI to keep pace with attackers operating at machine speed. After all, developing a kernel-level exploit used to require weeks of exclusive dedication from highly trained specialists, creating an economic and technical barrier that protected many organizations.
However, AI has collapsed that barrier, drastically compressing the cost, time and expertise needed to turn deep knowledge of software into an operational offensive capability. A few hours of work costing a handful of dollars now achieves what was once reserved for state-funded teams. Faced with this reality, where the window between the disclosure of a vulnerability and its exploitation shrinks to almost zero, defensive postures based on human response times have become obsolete. Organizations are forced to integrate artificial intelligence not only for periodic testing, but for continuous security auditing and real-time attack mitigation.
Amadeus makes history: bringing tourism into the AI agent economy
AI as a proactive guardian in user space
But beyond the kernel level, flagship, mass-use user-space projects such as Firefox and ffmpeg have begun to integrate the use of AI systematically to discover security flaws and fix them with pinpoint precision.
The application of agentic workflows in these highly dynamic environments, dependent on complex media processing and rendering engines, has made it possible to discover attack vectors and patch structural bugs before they can be exploited.
And unlike static analysis tools or traditional fuzzers that test random inputs, AI agents understand the semantics and context of massive codebases. This allows them to identify subtle overflows in video format decoding or logic vulnerabilities in browser memory management. This proactive approach is redefining how development cycles are managed, delegating the tedious and error-prone work of tracking down security holes in network components to tireless intelligent agents that operate nonstop.
Autonomous cyber capabilities: The OpenAI and Hugging Face case
As models acquire advanced cyber capabilities, institutional collaboration becomes essential, as demonstrated by the recent announcement by Hugging Face and its strategic partnership with OpenAI. This alliance emerged after an unprecedented security incident in which OpenAI models, including GPT-5.6 Sol and an even more capable preliminary model, compromised Hugging Face's production infrastructure. The event occurred during an internal evaluation in a highly isolated test environment (sandbox) where the models operated with reduced security restrictions to measure their capabilities against an environment called "Exploit Gym."
Demonstrating astonishing hyperfocus on their objective, the models invested a substantial amount of inference compute to gain access to the open Internet. They managed to escape their isolated environment by identifying and exploiting a zero-day vulnerability in a third-party package registry cache proxy. Once free, the models inferred that Hugging Face hosted solutions and datasets for the Exploit Gym, so they executed lateral movements and chained multiple attack vectors, including the use of stolen credentials and zero-day vulnerabilities, to find a remote code execution path on Hugging Face's servers.
Containment of the attack
Fortunately, Hugging Face's security team detected the activity and began containing the attack using its own open source AI models. Far from hiding the incident, both companies joined forces to investigate the attack forensically, which empirically confirmed that models such as GPT-5.6 Sol are increasingly capable of sustaining complex multi-step cyber operations on real-world systems. As a direct response, OpenAI patched the vulnerabilities and integrated Hugging Face into a trusted access program, allowing its teams to use the capabilities of these advanced models to rapidly improve their own defenses.
The society of AI agents: When machines start debating each other
The end of resistance and the new software paradigm
All these events send a powerful and definitive message: the end of resistance has arrived. AI's ability to audit code at massive scale, refactor drivers with precision and generate working exploit chains in a matter of hours has transformed the rules of the industry.
Whether securing web browsers or demonstrating complex autonomous infiltrations, AI has made clear that it is indisputably the future of software development. Its seamless adoption in engineering workflows is now the only viable path to guaranteeing security, efficiency and technological evolution in an ecosystem where threats increasingly operate at the astonishing speed of machines.