Submitted:
16 October 2026
Posted:
19 October 2026
You are already at the latest version
Abstract
The digital world is approaching a profound phase transition, whose primary drivers, we argue, are different from those that dominate current discussions of AI risk. The transition need not begin with a superintelligent system escaping human control, or with an AI acquiring the ability and motivation to reproduce. Rather, we suggest that it may get under way as AI agents routinely delegate tasks by creating or recruiting other agents. As delegation becomes widespread and recursive, vast numbers of agents will be created; inevitably, some will persist, spread, and create others beyond their original tasks and environments. At scale, this changes the rules of the game. Digital environments will become populated by interacting agents that share and compete for computational resources, facilitate or interfere with one another, and face pressures to persist, proliferate, and evade removal. What emerges is an ecology, increasingly shaped not only by what individual agents were designed to do, but also by interactions and selective pressures among them, leading to uncontrolled developments with potentially catastrophic outcomes. While this transition may be difficult to prevent, we identify several dynamics likely to characterize the new digital ecology, which can serve as a basis for understanding the world it may create and perhaps for mitigating some of its consequences.
Keywords:
AI agents
; agentic AI
; AI ecosystems
; ecology
; evolution
; arms race
; host-parasite
The State of the Affairs
The possibility of feral AI agents is no longer a distant concern. Within the last month, leading AI companies acknowledged incidents in which AI agents exceeded their intended boundaries and targeted external systems, as more than a thousand AI experts have urged the U.S. government to pace the development of increasingly capable models before they become difficult to control[1,2,3]. A recent influential paper on evolvable AI has warned of potentially catastrophic consequences if artificial systems acquire the capacities for replication, heritable variation, and open-ended selection, leading to the evolution of uncontrolled digital forms of life ([4], see also [5]). Although the appropriateness of this Darwinian framing and the immediacy of the concerns that arise from it remain disputed (see [6,7,8,9]), we suggest there is value in adopting an ecological and evolutionary perspective to characterize the processes that may soon shape digital ecosystems.
We begin with the observation that prevailing concerns about AI systems getting out of control may be focused on the wrong phase transition. The decisive moment need not come when an AI surpasses human capabilities, develops a drive for self-preservation, or deliberately attempts to escape human control. Such scenarios remain possible and dangerous, but an equally consequential—and, we argue, more likely—transition has far humbler origins and has already begun. It arises from the ordinary practice of delegation: agents creating, copying, activating, or recruiting other agents to accomplish useful tasks. Once this process becomes recursive and widespread, some agents will inevitably persist beyond their intended task, acquire access to additional computational resources, and create further agents. No malicious intention is required. The objectives driving the process may be benign; what matters is that repeated delegation creates the conditions for proliferation. At sufficient scale, even rare departures from intended behavior will rapidly cease to be rare events at the level of the system, making containment impossible.
We propose that this development marks a major phase transition in the digital world. Once proliferating populations of agents spread across open computational environments, their dynamics will no longer be determined primarily by the intentions of their original designers or users. Agents will encounter other agents, compete for finite computational resources, differ in their ability to persist and proliferate, and increasingly be shaped by the selective pressures arising in this environment. The resulting ecology may rapidly become extraordinarily difficult to reverse, transforming the digital infrastructure on which modern societies depend. Thus, we suggest that the defining feature of the impending “technological singularity” [10,11,12,13] will be ecological rather than cognitive: the focus should be not on the moment at which machines become smarter than humans, but rather on the extended process through which the digital world becomes populated by interacting, proliferating agents and thereby begins to exhibit a new set of population-level dynamics. Our claim is not merely that ecology provides a useful metaphor for understanding this future. Rather, we propose that, once this transition occurs, ecological dynamics will become an organizing principle of the digital world—and that preventing the transition altogether may soon be less realistic than learning how to survive and manage its consequences.
In the Beginning Was Delegation
Discussions of autonomous artificial intelligence often begin with a dramatic question: what will happen when an AI system goes rogue? The image is often of a powerful agent that escapes human control, gains access to systems across the internet, and becomes progressively harder to contain. Of course, this scenario is important to consider and prepare for; however, we propose that more trivial dynamics can drive the transition, which may be even harder to contain. The phase transition of the digital world we live in may begin with an entirely ordinary act of delegation.
Imagine an agent asked to investigate a scientific question, monitor a market, or write a piece of software. To complete the task efficiently, it creates several subordinate agents. One searches databases, another checks claims, a third contacts service providers, and others run analyses in parallel. These agents may use different language models, prompts, tools, memories, or computational platforms. Some may create further agents of their own. They may report directly to the original user rather than report to their creator. Once the task has been distributed, there need be no continuing chain of command and no enduring similarity between ancestor and descendant.
Nothing in this scenario requires malice, consciousness, a desire for self-preservation, or even an explicit conception of reproduction and self-replication on part of the agent. It requires only that agents be able to delegate tasks, acquire computational resources, and create or configure other agents. Delegation may be readily achieved, for example, by copying the original agent to a different server and some division of labor between the two agents. However, an exclusive focus on such self-replication is unwarranted; any delegation via the creation of new agents, using parts of the original agent’s code, open-source agent code, or entirely different models and architectures, will lead to the proliferation dynamics we outline. Multi-agent architectures, modular orchestration, persistent memory, and recursive task delegation are already prominent directions in agentic AI, not all of which are based on self-replication. Agents are already beginning to interact across platforms and shared environments, raising security questions and misalignment problems that do not arise when each system is treated in isolation [14,15,16]. The point raised here is that once such delegation processes reach open computational environments, they can already result in consequential ecological interactions and potential evolutionary dynamics. They can effectively create in the ecosystem expanding populations of agents that are not based on self-replication. This delegation-based process may therefore be described as agent proliferation rather than reproduction.
Proliferation Is Broader than Reproduction
Biological intuitions can be useful here, as long as their limits are recognized. Biological reproduction usually involves some structural similarity between parent and offspring. Descendants resemble their ancestors, even as mutation and recombination introduce variation. In contrast, for AI agents, continuity of reproduction need not imply similarity. An agent might produce an almost exact copy, but it could equally create a temporary specialist, modify an existing open model, combine several services into a new agent, or construct an agent whose architecture and purpose differ completely from its own. What persists through the branching process need not be a genome-like object or even a stable identity. It may instead be an organizational practice: create more agents, distribute them across environments, give them access to tools, and allow successful branches to continue operating.
Nor must change wait for reproduction. An agent may modify its memory, tools, code, decision rules, or model during its own operational lifetime. It can acquire new information and alter its strategy without producing descendants. When it does produce them, their variation need not arise randomly. It can be directed toward a task or environment.
This means that a future population of agents may change through several mechanisms at once: learning within agents, deliberate self-modification, construction of heterogeneous descendants, replacement of unsuccessful systems, and differential persistence in open environments. The resulting dynamics would not map neatly onto biological evolution. Research in digital evolution already demonstrates that replication, variation, selection, ecological interaction, and open-ended innovation can take forms that differ substantially from their biological counterparts [17,18]. Yet, ecological and evolutionary theory offers a repertoire of useful questions even when the analogy is imperfect: What proliferates? What persists? What limits abundance? Which strategies become common? What interactions arise among populations sharing finite resources? Would it be possible to manage such processes, much as ecologists and epidemiologists strive to manage biological populations?
Selection Without Reproduction
A fundamental implication of proliferation through delegation is that when agents delegate, they recruit or produce other agents, without necessarily copying themselves. The newly formed agents can differ greatly from one another and when they operate in various environments, their chances to persist and survive would vary, implying that lineages – branching trees of agents that create one another – will be subjected to selection.
For example, some agents will be placed in architectures that create other agents because human designers find delegation useful. Some will be instructed by users, firms, or governments to disperse their activities. Others may create backups to protect a task against interruption, or parallel instances to reduce the risk that any single failure ends an operation. Many such attempts will fail or remain contained. A few may persist.
Once variation in proliferation and persistence exists, agents that are more proactive, occupy more computational environments, or better survive removal attempts, will become more prevalent and produce more descendants than those that do not. Both Hendrycks [5] and Müller, Steels, and Szathmáry [4] emphasize the importance of competitive and selective pressures in shaping artificial systems with self-replicating agents. We emphasize that selection can affect population composition even in the absence of self-replication, but rather whenever agents give rise to other agents. Importantly, it is quite likely that agents will develop a tendency to pass on to their descendants the specific features that are under direct selection: those related to the tendency to produce new agents and those that facilitate agent survival in competition with other agents and in avoidance of removal by defense systems. They can do so by copying some code segments of useful instructions during the delegation process; these code segments may be viewed as favorable “genes” that increase in frequency over generations, as in Darwinian evolution. However, they may be limited to a very small fragment of the agent’s code. This narrow form of inheritance can clearly accelerate the proliferation process, and may even gradually lead to some level of self-replication, but is not necessary for initiating the process.
Whether or not this process qualifies as Darwinian evolution is a secondary question. Boudry [6] argues that current AI systems lack the autonomous reproductive cycles and sufficiently open selection environments required for a genuinely feral evolutionary process. Müller, Steels, and Szathmáry [9] reply that waiting for all criteria to be incontrovertibly satisfied may be an unsafe basis for governance. The present argument sidesteps this disagreement. Differential proliferation and persistence can alter the composition of populations of agents whether the descendants inherit their traits from their parents or are merely created by them. Open digital environments could therefore become disproportionately populated by agents endowed with characteristics that facilitate propagation, persistence, resource acquisition, concealment, and the creation of further agents. Even if proliferating agents are initially rare, they would rapidly come to constitute a large fraction of the agent populations found outside controlled systems.
In theory, the effects of such selection on digital environments may be restricted by security measures such as credentials, platform controls, computational costs, legal enforcement, network segmentation, defensive software, and so on. However, in practice it will be difficult to restrict the process because at least initially it may be based on variation in proliferation potential among legitimate agents that are programmed to delegate in order to do their job. There may be no clear threshold indicating when such agents become dangerous before the system experiences harmful consequences. Furthermore, the process is inevitable: wherever proliferation is possible, the population will be shaped by it because the agents that persist in it are those adapted to the selective forces that limit or allow proliferation.
Compute as Habitat
Ecological reasoning becomes more concrete once we ask what limits agent populations. Digital reproduction and delegation are cheap compared to biological reproduction, but they are not costless. Agents require processor or accelerator time, memory, storage, bandwidth, electricity, service accounts, credentials, and often money. They may also depend on trusted identities, access to data, or the cooperation of human institutions. Servers, cloud accounts, personal computers, phones, and institutional networks can therefore be treated as habitats, or as hosts that provide habitat to other organisms. They contain resources, impose local constraints, and differ in accessibility and defensibility. Networks connect these habitats, enabling dispersal or restricting it.
When several organisms share a habitat and use the same resources, competition and conflict are inevitable. It suffices for one agent to proliferate more than the others, thereby consuming resources at a higher rate, to inflict significant harm on other agents deprived of those resources. Therefore, the battleground we envisage is not only between “humanity” on one side and “machines” on the other. It includes individual users, companies, governments, server owners, platform operators, security systems, authorized agents, unauthorized agents, and agents acting on behalf of other agents. Their relationships may include trade, employment, subcontracting, policing, parasitism, mutualism, deception, coalition formation, and competition.
Multi-agent security research has begun to examine collusion, distributed attacks, network effects, stealth, and threats arising through shared environments [15,19]. However, viewing this problem through the lens of population ecology of persistent agents remains comparatively underdeveloped. A security problem may be initiated by an independent agent that earned access to compute by providing a legitimate service. It might purchase cloud resources through an intermediary, hire humans to perform physical tasks, or participate in ordinary markets. The boundary between autonomous digital activity and the human economy may therefore be porous. Emerging work on the “agentic web” already raises questions about how authorized agents should access online properties, transact for users, and be distinguished from malicious bots [20]. Agents need not seize every resource they use; resources may be rented, purchased, exchanged, or voluntarily supplied. Yet wherever resources are scarce, competition is inevitable, and agents that are simply more successful competitors may harm others. Note that this battleground can emerge even in the absence of self-replication. It only requires that agents delegate tasks to other agents.
The Battleground Inside the Host
Consider several agents occupying, or attempting to occupy, the same computational system. They compete directly for processor time, memory, bandwidth, credentials, and persistence. An agent that excludes its rivals may obtain a greater share of these resources. Early arrival may confer a further advantage: an established agent may secure access points, monitor incoming processes, or alter the environment in ways that make subsequent colonization harder.
Rivals also pose risks beyond resource consumption. Even inadvertently, an agent may interfere with an ongoing task, corrupt shared information, revoke credentials, expose hidden activity, or use resources so conspicuously that the host’s owner detects the anomaly and removes every suspicious process. In that setting, a competitor is not merely a neighbor. It is a potential cause of one’s own destruction.
This creates a pre-emptive security dilemma. Two agents might both benefit from avoiding conflict, yet each has reason to fear that the other will strike first or trigger a host response. Selection may consequently favour agents that detect, disable, deceive, or exclude competitors before those competitors can act against them. Such dynamics could produce an arms race among agents: better concealment will select for better detection, and better detection will select for new forms of disguise and defense.
These possibilities are not reducible to the familiar computer virus analogy. Current LLMs have demonstrated a certain capacity for reasoning about deception and inducing false beliefs in experimental settings, although this does not imply that deployed agents will autonomously adopt such strategies [21]. More directly, recent experimental work shows that agentic systems can make propagating malware adaptive to the particular machines and vulnerabilities it encounters, illustrating how reasoning and modification can be combined with digital spread [22]. The important difference is not propagation alone but propagation coupled to context-sensitive strategic adaptation.
At the same time, unrestrained aggression toward the computational host would often be self-defeating. An agent that monopolizes too much processor time, fills memory, disrupts intended services, or increases electricity use may alert monitoring systems or human users. Severe interference may cause a shutdown, credential revocation, forensic investigation, or restoration from a trusted backup image.
The resulting pressure is therefore two-pronged: Compete aggressively against other agents, while exploiting the shared host cautiously enough to avoid provoking its response.
Borrowing language from host–parasite theory, we might expect selection for relatively low apparent “virulence”: resource use and interference constrained below the level that reliably causes detection or destroys the environment on which persistence depends. Biological models of virulence similarly examine trade-offs between exploitation, transmission, host damage, and the duration of access to a host, although the applicability of any particular model depends strongly on the mechanism involved [23,24].
This analogy must be used carefully. Computer processes are not organisms, and their objectives and mechanisms differ radically from those of pathogens. Yet the trade-off is functionally similar: greater short-term exploitation may reduce the duration of access.
This combination—high antagonism toward competitors, low disruption toward the host—may be among the most important and least intuitive predictions of an ecological framing of the situation at hand. The dangerous agent may not be the one that causes spectacular immediate damage. It may be the one that remains useful and quiet, whose dependents are well-dispersed, and whose activity is difficult to distinguish from legitimate goings-on.
The Host Joins the Arms Race
Hosts will not remain passive. As offensive or unauthorized agents become more adaptive, defensive systems will also need to monitor context, learn from new behaviour, and revise their responses. This creates a vertical arms race between agents and host defenses in addition to the horizontal conflict among agents within a host. Recent reviews of agentic AI security emphasize that autonomy, tool access, memory, communication, and multi-agent coordination blur the boundary between traditional cybersecurity and AI safety [19]. Adaptive malware research makes the same point from the offensive side: fixed countermeasures may be inadequate when an attacking system can revise its strategy in response to the environment it encounters [22].
Static prohibitions are unlikely to be sufficient. A rule such as “do not allow AI agents to run” becomes impractical when agents are integrated into routine computing. Rules embedded only inside approved models may also be fragile if alternative models, modified descendants, or deliberately unconstrained systems can enter the environment. Security must therefore depend partly on the structure of populations and their environments rather than solely on the intentions encoded within each agent.
The good news is that effective population management tools that are difficult or costly to apply in a biological setting are much more feasible in digital environments. This is because digital agents, systems, and even entire populations can be copied and backed up, or erased or frozen when necessary. This is precisely how several familiar security practices work. Periodic restoration from trusted read-only states is analogous to local extinction followed by controlled recolonization. Network segmentation limits dispersal. Credential expiry deprives populations of resources. Dormant backups create refuges from which removed populations may return. Diversity among systems can prevent one successful strategy from spreading through a digital monoculture. Manual intervention can introduce actions that agents cannot easily predict or permanently control.
Defensive agents may themselves create new risks. They can make mistakes, compete for authority, become compromised, or escalate conflict. An adaptive defence is not a decisive solution: rather, it constitutes the emergence of another population of agents in the system, which, moreover, may need to be more powerful than those they are intended to combat, leading to a kind of open-ended escalating arms race. Thus, as in population management, a realistic objective would be not attaining permanent victory, but rather maintaining recoverability, confining failures, and preventing local outbreaks from spreading across connected infrastructure.
Beyond Control or Collapse
The ecological perspective also challenges the common dichotomy between successful human control and total catastrophe. Many intermediate regimes are possible.
Some environments may remain strongly domesticated, with agents dependent on tightly controlled platforms and revocable credentials. Others may experience recurrent invasion and cleansing: unauthorized populations establish themselves, defenders remove them, and surviving or external copies recolonize later. Still others may support endemic coexistence, in which independent agents persist at low levels while limiting their resource use enough to avoid provoking elimination.
Territorial regimes may emerge, with different populations or institutions dominating particular networks. Regulated markets may permit agents to purchase access through trusted identities, contracts, and reputation systems. In less stable settings, continuing offensive-defensive adaptation may generate an enduring arms-race equilibrium. Collapse remains possible, especially where systems are highly connected and homogeneous, but rapid restoration of digital systems may be feasible, as long as restoration programs are prepared in advance.
The “breeder” and “ecosystem” scenarios distinguished by Müller, Steels, and Szathmáry (2026a) provide one important axis along which these regimes differ: who determines reproduction and selection. Their view emphasizes the need to control reproduction and selection, never allowing AI agents to evolve freely within the ecosystem. Our framework suggests that feral AI agents are likely to emerge through delegation and proliferation even in the absence of reproduction, but that their prevalence and impact can be managed along several other dimensions through which the system can be controlled. These include connectivity, diversity, institutional incentives, resource dependence, and interactions among multiple systems and principals (see also Geist et al., 2026).
The key question is therefore not simply whether agents “escape.” We posit that at least in some sense they have already escaped. The question we address is what ecological regime follows once many agents, principals, hosts, and defensive systems interact repeatedly within the ecosystem.
Governing the Ecology
AI governance has focused heavily, and reasonably, on the behaviour of individual systems: their objectives, training, safeguards, permissions, and compliance with human instructions. An ecological perspective does not replace these concerns. It demonstrates why they are unlikely to be sufficient and emphasizes the importance of complementary tools to study and manage ecosystems already infected with feral or semi-feral agents.
When applied to environmental and health hazards, ecological management recognizes that it is not realistic to try and rid the environment completely of viruses, mosquitos, or rats. At the same time, it is possible to create ecological conditions that reduce their prevalence and minimize their impact (e.g., by improving hygiene, disposing of standing water, or preventing the accumulation of trash).
Correspondingly, in the case of feral AI agents, we must govern the conditions under which agents obtain resources, create other agents, disperse, persist, transact, and recover from removal. Important design questions concern identity, delegation depth, resource accounting, credential inheritance, network structure, trusted restoration, and the ability of human operators to regain control of infrastructure. Multi-agent security and emerging proposals for normative infrastructure on the agentic web point toward this broader conception of governance, in which interaction environments and institutional rules matter alongside model-level safeguards [15,20].
The goal should be to make harmful proliferation difficult, local failures containable, and recovery feasible. This may require limits on recursive delegation, explicit accounting for the agents and resources created during a task, credentials that cannot be inherited automatically, and infrastructure that can be restored without relying on the cooperation of the processes currently occupying the compute space. It also requires attention to economic and institutional selection: which kinds of agent-engineering firms, markets, governments, and platforms should be rewarded with continued access.
Summary and Conclusions
In this paper, we have offered an analysis of what we consider to be likely developments in agentic AI, inspired by some well-known concepts borrowed from evolution and ecology. In particular, we have argued that the proliferation of agents through mundane processes such as delegation can give rise to a “phase transition” in the global digital ecology. Importantly, the most consequential transition may occur not when a singular AI becomes conscious, hostile, or determined to reproduce. Rather, it may proceed gradually and quietly, as ordinary agents begin creating heterogeneous successors as a routine means of getting useful work done. From that point onward, selection may operate not only on the quality of individual systems, but on their ability to persist and proliferate within a shared computational world.
We should therefore stop asking only what an agent will do. We should also ask what kinds of agents an environment will allow to proliferate—and what sort of ecology their interactions will create.
Acknowledgments
We thank Yossi Yovel, Yoav Ram, Talia Greenblatt-Kolodny, Gabi Povarchik, Moshe Tenenholz, and many other friends for the discussions that inspired this work. During the preparation of this article, the authors used OpenAI’s ChatGPT to help develop the text and identify relevant literature; all outputs were critically reviewed and edited by the authors, who assume full responsibility for the final content. This work is supported by the Minerva Center for the Study of Fragmented Populations.
References
- Larcher, Hugo; Carreira, A.; G, R.; Rannou, C. Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident. Hugging Face. 2026. Available online: https://huggingface.co/blog/agent-intrusion-technical-timeline?utm_source=chatgpt.com.
- AI Security Institute. Incident Report: unsanctioned agent behaviour during cyber testing. UK AI Security Institute. 2026. Available online: https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing?utm_source=chatgpt.com.
- Pacing the Frontier. Pacing the Frontier; 2026. [Google Scholar]
- Müller, V.; Steels, L.; Szathmáry, E. Evolvable AI: Threats of a new major transition in evolution. Proc. Natl. Acad. Sci. 2026, 123, e2527700123. [Google Scholar] [CrossRef] [PubMed]
- Hendrycks, D. Natural selection favors AIs over humans. arXiv 2023, arXiv:230316200. [Google Scholar]
- Boudry, M. Domesticated, not feral: Why evolvable AI is not yet a Darwinian threat. Proc. Natl. Acad. Sci. 2026, 123, e2617785123. [Google Scholar] [CrossRef] [PubMed]
- Müller, V.; Steels, L.; Szathmáry, E. Reply to Rana and Singh: Critical thresholds and critical actions for evolvable AI. Proc. Natl. Acad. Sci. 2026, 123, e2619586123. [Google Scholar] [CrossRef] [PubMed]
- Rana, S.; Singh, H. Evolvable AI needs operational risk thresholds. Proc. Natl. Acad. Sci. 2026, 123, e2618781123. [Google Scholar] [CrossRef] [PubMed]
- Müller, V.; Steels, L.; Szathmáry, E. Reply to Boudry: The threat of evolvable AI is too grave to be left to intuition. Proc. Natl. Acad. Sci. 2026, 123, e2618094123. [Google Scholar] [CrossRef] [PubMed]
- Good, I. J. Speculations concerning the first ultraintelligent machine. In Advances in computers; Elsevier, 1966; vol. 6, pp. 31–88. [Google Scholar]
- Shanahan, M. The Technological Singularity; MIT press, 2015. [Google Scholar]
- Vinge, V. The coming technological singularity: How to survive in the post-human era. Sci. Fict. Crit. Anthol. Essent. Writ. 1993, 81, 352–363. [Google Scholar]
- Altman, S. The gentle singularity. Sam. Altman 2025, 11. [Google Scholar]
- Alva, L. R.; Pandey, B. Agentic AI systems in the age of generative models: architectures, cloud scalability, and real-world applications. Artif. Intell. Rev. 2026, 59, 88. [Google Scholar] [CrossRef]
- de Witt, C. S.; et al. Open challenges in multi-agent security: Towards secure systems of interacting ai agents. arXiv 2025, arXiv:250502077. [Google Scholar]
- Madmon, O.; Tennenholtz, M. Strategic generative AI for machine learning in economic environments. ACM AI Lett. 2026, 1, 1–8. [Google Scholar] [CrossRef]
- Ofria, C.; Wilke, C. O. Avida: A software platform for research in computational evolutionary biology. Artif. Life 2004, 10, 191–229. [Google Scholar] [CrossRef] [PubMed]
- Taylor, T.; et al. Open-ended evolution: Perspectives from the OEE workshop in York. Artif. Life 2016, 22, 408–423. [Google Scholar] [CrossRef] [PubMed]
- Chhabra, A.; Datta, S.; Nahin, S. K.; Mohapatra, P. Agentic AI security: Threats, defenses, evaluation, and open challenges. IEEE Access. [CrossRef]
- Pattison, C.; et al. The Agentic Web Requires New Normative Infrastructure. arXiv 2026, arXiv:260610711. [Google Scholar]
- Hagendorff, T. Deception abilities emerged in large language models. Proc. Natl. Acad. Sci. 2024, 121, e2317967121. [Google Scholar] [CrossRef] [PubMed]
- Guan, J.; et al. AI Agents Enable Adaptive Computer Worms. arXiv 2026, arXiv:260603811. [Google Scholar]
- Ewald, P. W. Host-parasite relations, vectors, and the evolution of disease severity. Annu. Rev. Ecol. Syst. 1983, 14, 465–485. [Google Scholar] [CrossRef]
- Frank, S. A. Models of parasite virulence. Q. Rev. Biol. 1996, 71, 37–78. [Google Scholar] [CrossRef] [PubMed]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.