The internet we knew is dead. The network built over decades on a fragile balance of human-written code, manual patches and perimeter defenses is being dismantled and rebuilt from the ground up. The catalyst for this structural change has a name of its own: Anthropic. With the commercial launch of Claude Opus 4.8 and the recent geopolitical opening of Project Glasswing, which has brought the feared Claude Mythos Preview model to 15 countries, including Spain, the digital ecosystem has crossed the point of no return.
The human internet after Anthropic
Previously, the digital ecosystem moved at human pace. Corporations and governments designed their online platforms as stone castles: static infrastructure was built, software was programmed line by line, and engineering teams watched their screens waiting for alerts. If a problem appeared, a committee of technicians met, analyzed the failure for days or weeks, and programmed a manual patch. Security was based on the perimeter (a good firewall at the entrance) and on obscurity (if no one discovered the flaw, the system was secure). The time factor favored the defenders.
Today, that world no longer exists. The digital territory post-Anthropic is dominated by machine speed. Online infrastructure can no longer be a static castle; it has to be a liquid organism that mutates and self-repairs in milliseconds. There is no longer time for a human to analyze code or approve a change on a server; systems defend and rewrite themselves in real time so as not to be devoured by armies of autonomous subagents attacking in parallel. IT maintenance has gone from being an office profession to becoming a war of asynchronous algorithms in which organizations that do not automate their structure 100% disappear from the network within hours.
The Mythos paradox: The end of digital secrecy
Until just a few months ago, cybersecurity was based on obscurity: if no one found a flaw in your code, you were safe. Claude Mythos has shattered that illusion at a stroke. In its first ultra-restricted deployment phase, limited to just 50 strategic partners in the United States, including Apple, Microsoft and Google, the model began to uncover critical vulnerabilities that had been embedded for decades in the planet's most sacred software.
The examination unearthed, for example, a 27-year-old flaw in OpenBSD and a hole in FFmpeg that had withstood more than 5 million automated attacks without ever being detected by a human eye. To grasp the scale of the impact, it is worth understanding what both of them do.
OpenBSD is the blueprint on which the internet's bunkers are designed. It is not a commercial company, but the free operating system focused on ultra-strict security that governments, banks and military services use to protect their networks. It was considered the most secure system on the planet; the fact that an AI found a crack hidden for almost three decades shows that no human code is infallible.
More than 23,000 potential flaws in open source software
FFmpeg is the invisible pipe through which all the world's video travels. It is the universal software suite that translates and plays multimedia files on platforms such as YouTube, Netflix or social networks. A hole there means an attacker could take control of a server just by playing a malicious video. The fact that this code had withstood millions of traditional cyberattacks and fell to the AI is evidence of a quantum leap.
Today, the Project Glasswing dashboard shows a metric that sends a shudder through the industry. The model has already located more than 23,000 potential flaws in open source software and critical databases around the world. The real bottleneck of the digital era is no longer finding computer errors; it is that human capabilities are completely overwhelmed when it comes to verifying, coordinating and applying patches at the frenetic pace at which the AI discovers them.
Faced with this defensive saturation, Anthropic has moved quickly. The firm has decided to expand access in a controlled way to 150 new organizations distributed across 15 non-Anglophone countries. The goal is to immediately bring in the essential sectors that were left out of the initial cohort and that sustain society's day-to-day life. Such as energy networks, drinking water supply, global telecommunications, healthcare systems and the main hardware manufacturers.
Spain's mission: strict confidentiality
Spain's inclusion in this global expansion is a sovereign defense measure. The Ministry for Digital Transformation and the Civil Service formally asked Anthropic to take part in testing Claude Mythos Preview in order to consolidate the country's influence in frontier AI governance and, above all, to protect its critical infrastructure.
Anthropic has stated that the expansion of the program is aimed at critical infrastructure operators, technology providers and entities whose activity is essential for millions of people. In Spain, participation takes place under strict confidentiality agreements and with the aim of strengthening the security of systems considered strategic for the economy and public services.
The rollout is structured around several priority areas. The first is the protection of sensitive environments, where information on vulnerabilities, audits and remediation processes remains within controlled infrastructure. The second is the review of legacy systems, especially in sectors such as banking, energy, telecommunications or public administration, where applications developed decades ago are still running and require progressive modernization. The third is accelerating vulnerability detection and correction processes through AI tools capable of helping to identify flaws, propose patches and support the validation and deployment of security solutions.
From employee to agent supervisor
The goal is to reduce the time between the discovery of a vulnerability and its effective correction, a factor that has become one of the main challenges of modern cybersecurity.
While the Mythos variant redefines defense at the state level, Claude Opus 4.8 is carrying out a silent revolution within the business fabric and commercial servers around the world. The paradigm of that passive chat in which a human asks questions has become obsolete. The introduction of the dynamic workflows feature has turned the AI into an autonomous conductor, where the model no longer executes code line by line, but instead plans a macro objective, spins up, coordinates and executes hundreds of subagents in parallel within a single session, and audits the results independently before reporting to the user.
In practice, this makes it possible to carry out massive migrations and restructuring of databases of hundreds of thousands of lines of code from start to final merge in the blink of an eye. The role of the knowledge worker has therefore shifted drastically. The human is no longer the executor, but the supervisor of a digital army.
This transition of the workforce has caused seismic changes in the daily structure and financial logic of companies.
Liquid online structures
For decades, corporate online platforms were designed under rigid schemes, with fixed servers, APIs with immovable permissions and sequential workflows. The arrival of autonomous agents is blowing up these software architectures. Static web structures are collapsing because they were not designed to handle the speed or volume of decisions made by an artificial intelligence.
The technical key to this change lies in the latest update to Anthropic's Messages API, which now allows system instructions to be injected directly into the message array in real time. Previously, if a company wanted to change the rules of the game or the context of a task for the AI, it had to restart the request, interrupting the workflow and destroying the message cache, which drove up costs and slowed down the process.
Now, corporate platforms operate as liquid networks. A corporation can update an agent's access permissions, modify its token budgets or alter the environment context in the middle of executing a task, without stopping the process and without the need for human intervention. Online infrastructure must become dynamic environments capable of negotiating authorizations, data traffic rerouting and logic flows in milliseconds, adapting to the unpredictable behavior of hundreds of subagents working in parallel.
The death of the man-hour
Throughout history, business resource accounting and planning have been governed by time. How many hours an employee works, how long a team takes to develop a project or how many consulting hours are billed. Opus 4.8 has destroyed the man-hour metric, replacing it with effort budget management.
With the new effort control built in from the factory, companies no longer pay for the "time" it takes the machine to process something, but for the cognitive depth required of it. At low effort, the model prioritizes response speed and resource savings, ideal for mechanical tasks or first-level customer service flows. At high effort, the AI activates self-reflection processes, chains subagents and thinks more deeply before issuing a response. It is the recommended level for complex code audits, reverse engineering or asynchronous flows that last hours or days without supervision.
The CFOs of multinationals no longer manage payrolls based on the attendance time of a programmer or an analyst; they now design budgets based on algorithmic complexity. The corporate currency is the token and the level of compute injected into long-running asynchronous flows.
The honesty index
An endemic problem that was holding back the full adoption of AI in companies' critical workflows was "confident hallucination": models that made serious errors in code but stated with total certainty that they had done the work perfectly.
Opus 4.8 introduces a radical change in its alignment focused on analytical honesty. The technical data show that the model is approximately four times less likely than its predecessor, Opus 4.7, to let flaws go unnoticed in the software it writes. The model now prefers to stop, flag its own uncertainties and warn the human supervisor about the lack of evidence before jumping to a conclusion.
This increase in reliability changes the design of human teams, because the work of a senior engineer no longer consists of hunting for the AI's hidden errors, but of validating the uncertainty warnings that the AI itself generates. Human oversight becomes strategic and high-governance, raising the bar for the qualifications demanded in modern offices. Junior profiles lose traction, while "agent directors" and "context auditors" become the most sought-after positions on the global market.
Cybercrime
The picture is not idyllic. The joint report by Anthropic and Verizon paints a reality in which attackers no longer need to be computer geniuses. 56% of the attackers detected in the past year already operate at medium-high risk levels, applying AI not to send phishing emails, which have fallen by 8.6%, but for complex operational tasks once inside the system, such as lateral movement or privilege escalation. Traditional security metrics for measuring an attacker's risk, such as the number of techniques they use, have become obsolete. Today, the most dangerous criminals simply design "agent architectures" that attack autonomously with minimal human intervention.
The epilogue of this transformation paints a radically new picture: the post-Anthropic ecosystem is, by definition, an Absolute Zero Trust scenario. Under the old computing paradigm, corporate networks had a clear border; once a user or a program passed the entry credentials, it was granted an implicit vote of confidence to move around the system. Today, that trust is operational suicide. In an environment where artificial intelligences can imitate human behavior to perfection and take control of systems from within, global corporations have had to transform themselves into genuine armored digital hives.
The time factor
In these hives, online infrastructure can no longer depend on the oversight of a systems administrator checking a dashboard in the morning. Today's corporate networks are forced to self-manage, mutate their configurations and self-repair constantly and autonomously, without interruption. This metamorphosis is the only path to survival against the relentless waves of scanning carried out by autonomous agency models 24 hours a day. If an online infrastructure remains static or takes more than a few minutes to react to an anomaly, automated subagents detect the crack, escalate privileges and exfiltrate the information before human security teams have poured their first cup of coffee.
Mythos, the Anthropic AI that has raised alarms over its destructive speed
This extreme urgency explains the speed with which the Spanish government and the other 14 countries selected for Project Glasswing are mobilizing their resources. This is not a simple software update; it is a race against the clock with a critically short expiration date.
Anthropic's own leadership has said that the current technological advantage is ephemeral. The firm's analysts estimate that the defensive exclusivity window will last, at best, between 6 and 12 months. That is the estimated time for competing labs, rival technology corporations or intelligence services of state actors such as Russia or China to independently replicate the reverse engineering and capability alignment of the Mythos class.
Before the end of 2026
The real danger is not that the technology exists, but its distribution. While Anthropic keeps Mythos under a strict digital embargo and only releases it to allied governments and critical infrastructure through the Cyber Verification Program, models developed by rival powers or organized cybercrime networks will leak onto the Dark Web black market without any kind of firewall, ethical limitation or security safeguard. When that military-grade attack AI becomes accessible to any criminal group on the planet, the nature of attacks on the network will change forever.
For this reason, Spain's strategy allows no delays. The country's critical infrastructure, from the plants that distribute energy to the nodes that support 5G or hospital systems, must be fully audited, secured and rewritten in secure code before this year ends. The defensive advantage against the next great wave of digital criminality can no longer be built with laws, committees or best-practice manuals; it must be built at machine scale, using autonomous software to defend the world before autonomous software is used to destroy it.




