Part Five of a Series on the Coming Labour Disruption
The Agentic Threshold: What Claude Mythos Tells Us About What Is Coming
A Threshold Has Been Crossed
On April 7, 2026, Anthropic made an announcement that should have commanded front pages across the world. It did not. The company stated, in technical language carefully chosen to be precise without being inflammatory, that its new model, Claude Mythos Preview, had “already identified high-severity vulnerabilities including some in every major operating system and web browser”—after a single user prompt, with minimal human direction (BBC, 2026). Anthropic further warned that “given the rapid advancements in AI, it won’t be long before such capabilities spread, potentially reaching individuals who may not use them responsibly” (BBC, 2026).
The company simultaneously decided not to release Mythos to the public. Instead, it has been made available to approximately 40 carefully selected organizations—banks, defence contractors, infrastructure operators—under strict controls, at significant commercial cost to Anthropic itself (YouTube/CSIS, 2026). Gregory Allen, senior adviser at the CSIS Wadhwani AI Center, estimated Anthropic is accepting a 6-to-8-month business disadvantage by withholding the model, because competitors are not showing the same restraint (YouTube/CSIS, 2026). Anthropic’s own draft materials acknowledge that Mythos “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders” (LinkedIn/Kanungo, 2026).
These are not marketing claims. They are safety warnings from the organization that built the system and understands it most completely. They deserve to be understood as such.
What Mythos Actually Does
To understand the labour and organizational implications of Mythos, it is necessary to first understand what distinguishes it from its predecessors—and why the distinction matters.
Every AI model that has preceded Mythos in public deployment has been, at its operational core, reactive. It responds to a prompt. A human identifies a task, frames a question, initiates the interaction. The AI performs a cognitive step. The human reviews the output, decides what to do next, and initiates the following step. The human remains the orchestrating agent throughout. The AI is a powerful tool—but tools require tool-users.
Mythos operates differently. It is, in the terminology researchers use, a fully agentic system at expert level. It can be given a goal—”find exploitable vulnerabilities in this codebase”—and it then autonomously plans the investigation, selects the methods, executes the analysis, tests hypotheses, adapts when approaches do not yield results, documents findings, and produces the equivalent of an expert penetration test report, without requiring human direction at any intermediate step (Bishop Fox, 2026). It does not wait to be told what to do next. It decides.
The cybersecurity domain was the first arena in which this capability was demonstrated at expert level because it is measurable, testable, and has clear success criteria: either the model finds the vulnerability or it does not. It does. In testing, Mythos outperformed all but the most elite human cybersecurity specialists—not on simplified benchmark tasks, but on real operating systems and web browsers in current deployment (Anthropic, 2026). The leaked internal materials describe the “Capybara” capability tier of Mythos as “dramatically” better than Claude Opus 4.6 at programming and reasoning tasks—the foundation on which agentic cybersecurity work rests (LinkedIn/Kanungo, 2026).
Why Cybersecurity Is the Leading Edge
The cybersecurity domain is not incidentally where Mythos has been deployed. It is specifically where agentic AI has been tested because it provides a controlled and measurable environment for capability assessment. But the cognitive capabilities required for expert-level autonomous cybersecurity work—sustained multi-step reasoning across complex systems, hypothesis generation and testing, adaptive problem-solving, pattern recognition in novel environments—are not unique to cybersecurity. They are the same capabilities required for expert-level work in legal analysis, financial modelling, strategic planning, medical diagnosis, scientific research, and organizational management.
What Mythos has demonstrated in the cybersecurity domain is a proof of concept for agentic expert performance that will diffuse. The specific domain is the leading edge; the underlying capability is general. When a system can autonomously reason at expert level across a complex, real-world problem with many interdependencies and no predetermined solution path, it has demonstrated something that transcends the specific application. The question is not whether those capabilities will be applied to other domains. It is how quickly, and by whom.
Workers and watchdogs were entirely absent from the process by which Mythos’s 40 initial partner organizations were selected, with no public consultation, no labour representation, and no disclosure of which sectors are now operating with access to these capabilities (Tech Policy Press, 2026). That absence is itself informative about whose interests are centered in the deployment decisions being made.
The Organizational Implications: The Middle Is Gone
For organizations, the deployment of agentic AI at Mythos-level capability does not represent incremental efficiency improvement. It represents the elimination of entire organizational layers that have existed for as long as organizations have been complex enough to require them.
The middle management layer in most organizations exists to perform a specific set of functions: translating strategic direction from senior leadership into operational tasks for front-line workers; gathering information from the front line and synthesizing it for senior leadership; coordinating across teams and functions; monitoring progress and managing exceptions; making routine decisions within established parameters. Gartner projects that 20% of organizations will use AI to flatten their hierarchies by end of 2026, eliminating more than 50% of current middle management positions (LinkedIn/Keith, 2026). Amazon has already cut 14,000 corporate roles, explicitly citing “AI enabling leaner structures.” Wall Street banks plan to eliminate approximately 200,000 roles over the next three to five years, heavily concentrated in middle-layer oversight functions (Kanungo, 2026).
A typical middle manager spends roughly 60% of their working week on four tasks: reporting up, coordinating laterally, reviewing down, and translating between organizational levels (Kanungo, 2026). AI agents can now assemble status reports in seconds by pulling directly from CRM systems, project management tools, and communications platforms. They can coordinate across teams by monitoring task progress and flagging exceptions without human intervention. They can review output for consistency with established parameters faster and more comprehensively than human managers. They can translate strategic objectives into operational tasks by decomposing goals autonomously. The middle manager’s week, examined task by task, has been automated. Not hypothetically. Currently (Kanungo, 2026).
MIT Sloan’s 2026 AI research shows that in companies already deploying agentic AI at scale, the span of control—reports per manager—has expanded from the historical norm of seven to as high as fifteen in some divisions (Kanungo, 2026). The organizational mathematics are straightforward: if one manager can now effectively oversee fifteen workers instead of seven, an organization needs roughly half as many managers as before. That is not projection. It is documented current data from organizations already in transition.
The Entry-Level Collapse: The Generation Without a Ladder
While the elimination of middle management is perhaps the most structurally visible consequence of agentic AI deployment, a quieter and in some ways more consequential shift is occurring at the entry level of white-collar careers.
Entry-level professional positions have historically served two functions: they provided organizations with low-cost labour for routine cognitive tasks, and they provided young workers with the experience, mentorship, and credentialing required to progress toward more senior roles. The first function is being eliminated by AI. The second function disappears with it.
Anthropic’s own survey of 81,000 Claude users found that people in early-career positions in highly AI-exposed occupations express the highest concern about economic displacement—and that those experiencing the largest productivity speedups from AI also express the highest anxiety about job security (Anthropic, 2026). The latter finding is counterintuitive and important: using AI to become more productive does not reduce the fear of being replaced by AI. It heightens it, because the worker can see precisely which of their tasks the AI is performing—and extrapolate what happens when it performs all of them without requiring the human wrapper.
Anthropic’s March 2026 labour market report, based on real enterprise usage data, found that 75% of programming tasks and 67% of data entry tasks are already being performed by Claude in observed enterprise deployments (The AI Corner, 2026). These are not tasks that were previously considered automatable—they were the preserve of educated, credentialed, early-career professionals. The entry-level programmer, the junior analyst, the graduate trainee accountant—the roles that constituted the first rung of professional career ladders—are being automated not at the level of possibility but at the level of current documented practice.
Anthropic CEO Dario Amodei stated in mid-2025 that half of entry-level white-collar jobs could vanish within one to five years (WhatLLM, 2026). Microsoft AI CEO Mustafa Suleyman went further, projecting that most white-collar tasks—not jobs, tasks—would be automated within twelve to eighteen months (WhatLLM, 2026). The distinction between tasks and jobs is less reassuring than it first appears: when every task that fills a working day is automated, the practical difference is academic.
The Speed Problem: This Is Not a Gradual Transition
Every previous wave of technological displacement in human history played out slowly enough that the people experiencing it could see it coming, could observe peers navigating it, could receive imperfect but real social and institutional support, and could make at least partially informed decisions about how to respond. The displacement was painful. It was often unjust. But it was, in most cases, gradual enough to be individually navigable.
The AI transition, and specifically the agentic AI transition that Mythos represents, does not share this characteristic. It is moving at the speed of software deployment, not the speed of industrial retooling or agricultural mechanization. An organization can go from human-intensive to AI-native in the time it takes to integrate an API and adjust its workflows. The competitive pressure to do so—documented in Parts One and Three of this series—operates on quarterly earnings cycles, not generational ones.
Anthropic’s own agentic coding trends report for 2026 shows that the adoption of AI agents for complex, autonomous work tasks is not spreading at a linear rate (Anthropic, 2026). It is accelerating. The organizations that have moved earliest are widening their productivity advantages over those that have not—creating the competitive pressure on laggards that Part One identified as the mechanism that makes loyalty pledges structurally impossible to honour. The transition is self-accelerating. The longer an organization waits, the more ground it loses, and the more desperate its eventual response.
For the individuals who are the subject of this transition—the mothers and fathers paying mortgages, the young people building careers, the middle-aged professionals who restructured their financial lives around the assumption that their cognitive skills would retain value—there is no warning period commensurate with the speed of change. By the time the displacement is visible in unemployment statistics, in sector contraction, in benefit claims, the transition will already be substantially complete. The data is already in the system. The layoffs already announced. The entry-level positions already eliminated. The middle management layers already scheduled for restructuring. Most of the people who will be displaced in the next three years do not yet know it.
The Tribalization Risk: The Social Fracture No One Is Naming
There is a consequence of mass rapid displacement that receives almost no attention in the mainstream analysis of AI’s labour market impact, and it may ultimately matter more than any economic metric. It is the social and political fracture that follows when large numbers of people lose their economic function simultaneously, without institutional support adequate to absorb the shock, and without a shared narrative that makes sense of what has happened to them.
Economic displacement does not produce passive resignation. It produces meaning-seeking. When the structures around which people have organized their lives—employment, professional identity, the expectation of progress, the sense of contributing to something larger than themselves—collapse, people do not accept the collapse. They look for explanations, for community, for enemies, and for alternatives that promise restoration of what was lost. This is not a pathological response. It is a predictable and historically documented human response to rapid structural disorientation.
The tribalization that this process produces—the fracturing of shared social reality into mutually hostile communities organized around incompatible explanations of what went wrong and who is responsible—is already visible. It preceded the AI displacement wave because austerity, inequality, and institutional failure had already been straining the social fabric for decades. AI-driven mass displacement does not initiate this process. It accelerates it, expands its scale, and removes the possibility of the gradual recovery that has historically allowed societies to reintegrate after disruption.
The AI utopians, who project a frictionless transition to abundance, offer a narrative that the evidence does not support and that the people experiencing displacement will not find credible. The AI dystopians, who project permanent and total catastrophe, make the error of assuming that human societies do not adapt—they do, eventually, though sometimes at horrific cost. The vast majority—the ostriches who proceed as though nothing fundamental is changing—are not being foolish. They are responding rationally to a situation that is genuinely difficult to look at directly.
What is needed, and what is not yet being produced at adequate scale, is honest, evidenced, unsentimental analysis that takes seriously both the scale of what is coming and the genuine human capacity—not institutional capacity, but individual human capacity—to navigate it. The people who will be most redundant in an AI-driven economy are not those who lack specific technical skills. They are those who have not developed the underlying human capabilities—analytical, adaptive, creative, resilient—that remain meaningful regardless of what the landscape looks like.
That is not a comfortable conclusion to offer to a person trying to pay a mortgage, manage a household, and make sense of a world changing faster than any previous generation has been asked to absorb. It is, however, the honest one. And it is the one that actually points toward something actionable—not for the institutions that are failing to respond at adequate speed, but for the individual who still has time, however narrowing, to invest in the capabilities that no transition can make obsolete.
A Note on Mythos and Control
Anthropic’s decision to withhold Mythos from public deployment deserves to be taken seriously as a signal, not merely as corporate caution. The company is absorbing a 6-to-8-month competitive disadvantage to maintain control over a system it has assessed as too dangerous for unrestricted deployment (YouTube/CSIS, 2026). That judgment—from the organization with the most complete knowledge of the system’s capabilities—should be weighted accordingly.
But the history of frontier AI models suggests that “controlled deployment” is a temporary state. Every previous model deemed too powerful for public release has eventually been released, either by its developers or by competitors who independently reached comparable capability. Chinese state-sponsored actors were already exploiting Claude’s agentic capabilities in November 2025 to breach targets globally, having found it “remarkably simple” to bypass existing safeguards (BBC, 2026). The 40 organizations currently with Mythos access include some of the world’s largest financial institutions and defence contractors—organizations with every incentive to deploy its capabilities in contexts far beyond the cybersecurity applications for which it was initially vetted.
The controlled deployment of Mythos is not a solution to the problems this series has traced. It is a delay. The capability exists. The competitive pressure to deploy it commercially exists. The institutional frameworks to govern its deployment do not yet exist and are not being built at the necessary speed. What comes after the delay is not a managed transition. It is a race.
The person who is the subject of this series—the individual trying to navigate a world being reorganized around them at a speed no institution can match—cannot wait for the race to conclude before deciding how to respond. The time for that decision is now, before the displacement curve steepens, before the entry-level positions are fully eliminated, before the middle management layer is gone, and before the economic ecosystem that sustains ordinary life has absorbed the full force of what Mythos and its successors will bring.
References
Anthropic. (2026, February 25). Labor market impacts of AI: A new measure and early evidence. https://www.anthropic.com/research/labor-market-impacts
Anthropic. (2026, April 21). What 81,000 people told us about the economics of AI. https://www.anthropic.com/research/81k-economics
Anthropic. (n.d.). Project Glasswing: Securing critical software for the AI era. https://www.anthropic.com/glasswing
Anthropic. (n.d.). Transparency hub. https://www.anthropic.com/transparency
BBC. (2026, April 17). What is Anthropic’s Claude Mythos and what risks does it pose? https://www.bbc.com/news/articles/crk1py1jgzko
Bishop Fox. (2026, April 13). Anthropic’s Claude Mythos Preview: The AI cybersecurity inflection point. https://bishopfox.com/blog/anthropics-claude-mythos-preview-the-ai-cybersecurity-inflection-point
Kanungo, S. (2026, April 20). Why 50% of middle managers could be replaced by AI by 2027—and what to do if you’re one. https://shawnkanungo.com/blog/why-50-of-middle-managers-could-be-replaced-by-ai-by-2027-and-what-to-do-if-youre-one
LinkedIn/Keith. (2026, February 2). AI to eliminate middle management positions by 2026. https://www.linkedin.com/posts/keithand_gartner-projects-that-20-of-organizations-activity-7424459592757288960-aeQY
LinkedIn/Kanungo, D. (2026, April 12). Anthropic’s AI model raises verification concerns. https://www.linkedin.com/posts/danielhulme_so-anthropic-announced-an-ai-model-too-powerful-activity-7449409984670515200-NP5D
Tech Policy Press. (2026, April 22). Anthropic warned big companies about Mythos. Workers and watchdogs need a seat at the table. https://techpolicy.press/anthropic-warned-big-companies-about-mythos-workers-and-watchdogs-need-a-seat-at-the-table
The AI Corner. (2026, March 5). Anthropic just showed us which jobs AI is actually replacing. https://www.the-ai-corner.com/p/anthropic-ai-jobs-report-2026
WhatLLM. (2026, February 15). The white-collar existential crisis: How AI killed the meaningless job. https://whatllm.org/blog/white-collar-existential-crisis-ai-agents
World Economic Forum. (2026, April 19). Anthropic’s Mythos moment: How frontier AI is redefining cybersecurity. https://www.weforum.org/stories/2026/04/anthropic-mythos-ai-cybersecurity/
YouTube/CSIS. (2026, April 16). Anthropic’s Mythos: What it is and what it is capable of. Center for Strategic and International Studies. https://www.youtube.com/watch?v=vQOROpbUdzA