In April 2026, when Anthropic announced that its latest model, Claude Mythos, had entered the preview stage, the mood across the global tech community was not the celebratory frenzy of the past, but something closer to the tense vigilance following a nuclear test.
Over the past three years, public understanding of artificial intelligence (AI) has largely remained at the level of a “very smart encyclopedia” or a “creative assistant that can write poetry.” However, the arrival of Mythos marks a clear turning point in AI’s evolution: we are moving from the era of generative AI into that of agentic AI. This is no longer merely an upgrade of a tech product, but the emergence of a digital entity capable of directly intervening in the real world, reshaping national security systems, and even challenging the very definition of human labor.
Too Powerful to Be Released to the Public?
Mythos is one of the latest models developed by Anthropic under its broader AI system known as Claude, which competes with OpenAI’s ChatGPT and Google’s Gemini.
In early April, Anthropic unveiled the model in the form of a “Mythos Preview.”
During internal red-team testing, Mythos demonstrated insights far beyond the reach of human experts. It identified a vulnerability in the security-focused OpenBSD system that had remained undetected for 27 years despite review by countless top engineers. More alarmingly, it discovered multiple high-risk vulnerabilities in the Linux kernel and was able to automatically “chain” them together, instantly generating a complete attack script capable of seizing full system control.
This capability fundamentally alters the underlying economics of cybersecurity. In the past, discovering a “zero-day vulnerability” required months of effort; with Mythos, it becomes a scalable, automated pipeline. This suggests that the firewalls, encryption protocols, and access control systems underpinning human digital civilization may now face structural risks of collapse at the technical level.
Anthropic CEO Dario Amodei has argued that Mythos is too risky for public release. This is not the first time an AI company has restricted access due to a model’s power. In 2019, well before the era of ChatGPT, OpenAI took similar measures with its GPT-2 model. Notably, Amodei was then a core researcher at OpenAI.
From National Security Threat to Strategic Asset
As Mythos’s capabilities came to light, technical debate quickly escalated into geopolitical competition. This is no longer just a Silicon Valley story—it has become a strategic concern for the White House and the Pentagon.
In February 2026, U.S. President Donald Trump ordered federal agencies to halt the use of Anthropic’s products, criticizing the company as being controlled by “left-wing lunatics.” The Pentagon even placed it on a “supply chain risk” blacklist, partly because the company refused to allow its AI to be used for all military purposes, including autonomous weapons. Yet the sheer power of Mythos forced a shift in political reality.
On April 17, Dario Amodei was invited to the White House for high-level meetings with senior advisors. Despite ongoing legal disputes over national security risks, the U.S. government clearly could not afford to lose access to Mythos. This paradoxical situation—litigating and cooperating at the same time—reveals a stark truth: AI has become the nuclear weapon of the digital age. When Washington seeks agreements on scaling technology, it is not merely promoting innovation, but pursuing absolute control over this new force.
AI is no longer just a technological product—it has become a strategic asset at the level of national security.
Imagine, for instance, if Mythos fell into the hands of hostile actors or transnational criminal organizations. It would no longer be an assistant, but a strategic weapon capable of crippling power grids, contaminating water systems, or freezing bank accounts—systems essential to everyday life.
Project Glasswing
To address the disruption caused by Mythos, Anthropic launched Project Glasswing. This initiative brings together 12 major tech companies—including Microsoft, Amazon, and Google—as well as government agencies, with the aim of transforming Mythos’s capabilities into defensive power.
In essence, Project Glasswing establishes a new kind of “digital gatekeeping” mechanism. A private company effectively determines which institutions and which countries’ infrastructure are eligible for protection by the most powerful AI systems. When technological advancement reaches a point where access itself becomes a matter of survival, technology ceases to be neutral.
This concentration of power raises serious ethical concerns: if only financial giants and tech empires can access Mythos’s protection, will small and medium-sized enterprises, developing nations, and ordinary individuals be left as vulnerable prey in a digital wilderness?
China’s “Lobster Farming” Craze
Before discussing how Mythos might threaten the world, it is worth turning to a more grounded example: a recently popular AI trend in China known as OpenClaw, colloquially referred to as “raising lobsters.”
OpenClaw is an open-source AI agent framework developed by Austrian developer Peter Steinberger. Technically, it is not a standalone AI model, but a framework that allows users to plug in different large language models as the “brain,” while connecting them to tools such as browsers, file systems, email, calendars, and messaging apps.
Users can issue commands via messaging platforms—such as checking news, tracking stocks, organizing data, or scheduling tasks—and OpenClaw can read on-screen content, interpret instructions, and execute actions across multiple systems.
This agent-based model transforms AI from a tool into something closer to an actor, marking a fundamental shift in how humans interact with computers. Traditional AI responds with words; OpenClaw acts. It can read screens, simulate mouse clicks, and jump between applications, performing tasks ranging from sending emails to generating complex financial reports—essentially functioning as an invisible digital worker.
A Digital Labor System
Local governments in China have begun to take note. Shenzhen’s Longgang District has proposed subsidies of up to 10 million yuan for “one-person companies,” where a single founder could theoretically operate a business using AI tools. Wuxi in Jiangsu Province has also offered funding to promote AI agent applications in robotics and industrial settings.
Against the backdrop of economic pressure and youth unemployment, some policymakers view AI agents as tools for boosting productivity. If AI can replace parts of administrative, customer service, or even research work, then a single individual might indeed be able to run a small enterprise. Tools like OpenClaw are increasingly seen as the foundation of an “individual productivity revolution.”
At the same time, they are reshaping the structure of work. Tasks traditionally performed by entry-level white-collar workers—handling emails, organizing data, drafting documents—may gradually be automated.
This shift may not immediately eliminate jobs, but it will redefine them. A research assistant may spend less time compiling data and more time verifying AI outputs; an administrative assistant may shift from scheduling meetings to overseeing automated workflows.
However, the benefits of this transition are unlikely to be evenly distributed. Large tech firms and capital-intensive enterprises are best positioned to leverage AI for efficiency gains, while roles reliant on routine administrative tasks may face greater pressure.
In other words, OpenClaw represents not just a new tool, but an emerging digital labor system.
A Shift in Power
For decades, computers have been seen as tools. Whether word processors or search engines, humans remained in control. AI agents are changing that relationship.
When a system can understand goals, break down tasks, and act autonomously, it begins to resemble a hired digital worker rather than a passive tool. Some commentators now describe these systems as “digital laborers.”
The “lobster farming” phenomenon reflects a deeper shift: people are beginning to see AI as a form of domesticated, accumulable labor. If labor no longer depends on human time or physical effort, but on computational resources, how should the value of work be measured?
More profoundly, human roles may shift toward extreme supervision. The widespread adoption of tools like OpenClaw hints at a new inequality: those with access to computational power and technical permissions can command thousands of “lobster workers,” while those relying on traditional skills may see the marginal value of their labor approach zero. This is not just about technological unemployment, but about redefining the very nature of human labor in the age of AI agents.
A New Normal for AI
The rise of OpenClaw may be only an early sign of broader changes. In future work models, humans may focus on setting goals and overseeing processes, while AI systems handle routine execution. This resembles industrial automation—but this time, it is not just physical labor being automated, but elements of cognition and decision-making.
If widely adopted, this model will not merely improve efficiency; it will fundamentally challenge how society defines “work.” When humans no longer need to directly complete tasks, the value of work may shift toward creativity, judgment, and responsibility. AI is not just replacing jobs—it is redefining what counts as work.
This transformation is mirrored in the emergence of Claude Mythos. If OpenClaw reshapes individual workflows, Mythos touches the structure of society itself. When AI can intervene in infrastructure, cybersecurity, and institutional operations, its impact extends beyond labor markets to the distribution of power and risk.
A Double-Edged Sword
AI is undeniably powerful, but it would be simplistic to view it purely as a threat. Both OpenClaw and Mythos highlight its potential.
OpenClaw can reduce operational barriers and automate repetitive tasks, freeing humans from mundane work. For businesses, AI agents can restructure productivity, allowing small teams to achieve what once required large workforces.
In cybersecurity, vulnerabilities often persist not because they are impossible to fix, but because they are never discovered. Identifying them requires expertise, time, and uncertainty. If models like Mythos can scan systems at scale, the cost of vulnerability discovery could drop dramatically, transforming cybersecurity from a scarcity-driven field into an automated process.
Real-world incidents illustrate the stakes. The Optus data breach affected around 9.5 million people, while the Medibank hack exposed sensitive medical data, some of which was leaked on the dark web. These cases show that cybersecurity is not just technical—it directly impacts privacy, identity, and social trust.
If Mythos can turn vulnerability discovery from a rare, reactive process into a proactive, continuous one, defenders could, in theory, match or even surpass attackers in speed and scale.
But therein lies the problem.
If both vulnerability discovery and exploitation become automated, hacking itself could be industrialized. Attacks that once required advanced expertise could become replicable and scalable. In the wrong hands, this would create systemic risks rather than isolated incidents.
Who Controls AI?
Ultimately, the debate sparked by Claude Mythos is not about the technology itself, but about who controls it—and who oversees its use.
If such models remain in the hands of a few corporations and governments, AI will become not just an economic resource, but a form of power. This power could reshape international relations, corporate competition, and social hierarchies.
In this process, AI is not the decisive factor—people are.
Technologies like Mythos and OpenClaw do not point to a single future. They can strengthen defense or enable attack, improve efficiency or deepen inequality, repair systems or break them. This duality may be the defining characteristic of AI.
The key is not whether AI is “good” or “bad,” but that it amplifies both possibilities.
The Fear of AI
Public fear of AI is not new. From early automation replacing blue-collar workers, to generative AI threatening white-collar jobs, to extreme scenarios of runaway superintelligence—these anxieties have accompanied every major technological shift.
While such fears are not unfounded, they have also become part of the tech industry’s narrative. One must ask: is “fear” being used as a marketing strategy?
In Silicon Valley, claiming that a model is too powerful to release often generates media attention and higher valuations. The rollout of GPT-2 in 2019 followed a similar pattern. By invoking apocalyptic scenarios, companies may also be reinforcing their own authority.
As former UK National Cyber Security Centre head Ciaran Martin has noted, while Mythos is impressive, many risks may be overstated. Most real-world cyberattacks still stem from basic human security failures. Overemphasizing AI threats can distract from improving digital literacy and social protections.
Technology Is Neither Good nor Evil
Whether Claude Mythos or OpenClaw, both illustrate the dual nature of AI: it is both a digital craftsman capable of repairing the internet’s foundations and an invisible hacker that could destabilize civilization.
The central argument is not to judge AI as good or bad, but to highlight a pressing reality: when technology evolves to influence national sovereignty and the nature of labor, who controls it matters more than the technology itself.
AI is not just technological evolution—it is a mirror reflecting existing fractures in power and weaknesses in defense. Facing systems like Mythos, the response should not be limited to restriction or competition, but should include building fair and transparent frameworks for access and accountability.
In the end, humanity’s fate will not be decided by algorithms hidden in sandboxes, but by whether we have the courage to redefine a shared boundary of safety in the midst of rapid technological acceleration.
The spark of AI has already been lit. Whether it becomes a guiding light or a wildfire depends not on the strength of the flame, but on how we choose to wield it.