Learning to Replicate: Anthropic's Self-Replication Research
Updated: Oct 23
From smart homes and digital media to enterprise networks and industrial control systems, new research raises important questions about autonomy, cybersecurity, and our connected world.
DEW Diligence, LLC | October 21, 2026 — DEW Diligence is not affiliated with Anthropic; this is independent commentary.
Key Takeaways
Anthropic's Frontier Red Team showed that a 26-billion-parameter model can run across 15 inexpensive Android TV boxes, that a 9-billion-parameter model can set up inference on 8 of 9 unfamiliar hardware types, and that a more capable model (GLM-5.2) can chain the full replication workflow in a simplified test network.
Anthropic projects that small models may be "months away" from this capability level. That forecast sits on a general capability trend line and on experiments that deliberately trivialized the hardest step of breaking into devices.
No uncontrolled worm has been demonstrated in homes, businesses, or industrial facilities. The experiments show capable components of autonomous replication under controlled conditions.
By Anthropic's estimates, such a replicator would be slow (one to two days per replication) and compute-poor (a botnet the size of Kimwolf would yield roughly 40 H100-equivalents of inference).
It is important to consider asset visibility, patching and hardening, network segmentation, strong identity controls, behavioral monitoring, and operational resilience. The challenge is applying them consistently across a growing population of connected devices.
Artificial Intelligence (AI): Entering a New Phase of Cybersecurity Risk
Imagine an ordinary evening at home. Your smartphone is connected to Wi-Fi, a smart television streams a movie, security cameras monitor the property, and a smart thermostat adjusts the temperature. Perhaps your door locks, alarm system, garage door, and lighting are also connected. Each device serves a different purpose, yet many depend on the same underlying network infrastructure, software, cloud services, and digital identities.
What if malicious software operating on one compromised device could use intelligence to identify other accessible devices, adapt to their operating environments, and potentially establish additional copies of itself?
On October 12, 2026, Anthropic's Frontier Red Team published AI models may be capable of self-replicating without access to a datacenter. Their research examined whether smaller AI models could independently propagate across distributed computing environments, including relatively inexpensive Internet of Things (IoT) hardware. The findings are intriguing, technically significant, and deserving of attention.
They are also important to interpret correctly. Anthropic has not demonstrated an uncontrolled AI worm spreading through real-world homes, businesses, or industrial facilities. Its experiments instead demonstrate capable components of autonomous replication under controlled conditions. Responsible cybersecurity begins with understanding demonstrated capabilities, credible risks, and the safeguards available today.
What Did Anthropic Actually Discover?
Anthropic investigated autonomous replication and adaptation (ARA). ARA is the potential for an AI system to establish and propagate instances of itself without continuous human direction. Traditional computer worms have existed for decades, spreading by exploiting vulnerable systems, reusing credentials, or abusing network services. Their behavior generally depends on predefined software instructions, although conventional malware can already incorporate sophisticated decision-making.
AI introduces an important difference, and that is the ability to reason through unfamiliar technical environments, troubleshoot failures, and develop new approaches through changing circumstances. Anthropic examined whether sufficiently capable models could perform these activities while also operating on comparatively limited computing resources. Anthropic describes three threat models. The first is a small model that replicates by compromising host devices and running directly on them (a "weight-based replicator"). The second is a large, advanced model focused on data centers and other high-compute targets. The third is an API-based replicator that propagates by stealing credentials for commercial inference services. The research examines only the first, which Anthropic considers the nearest-term and most directly testable.
Within that scope, three findings stand out.
First, meaningful AI inference does not necessarily require datacenter infrastructure.
Researchers distributed Gemma4-26B-A4B, a 26-billion-parameter mixture-of-experts model with roughly four billion active parameters, across 15 inexpensive off-brand Android TV boxes, each costing approximately $55 and equipped with 4GB of RAM. Working together, the devices performed inference at approximately 0.39 tokens per second. That is extremely slow compared with contemporary cloud-based AI services. Nevertheless, it demonstrated that inexpensive consumer hardware can collectively execute a model that would be impractical to run on an individual device.
Second, smaller AI models demonstrated considerable adaptability.
Qwen3.5-9B, a nine-billion-parameter model, successfully established an inference server across eight of nine tested hardware types, ranging from NVIDIA Jetson Nano boards to Mac desktops and RTX 4090 workstations. These environments included different computing architectures and configurations, illustrating how capable coding agents can work through compatibility and deployment challenges.
Third, a more capable model demonstrated reliable propagation in a simplified experimental network.
In a firewalled Google Cloud environment consisting of eight nodes running redroid, an Android 15 emulator, GLM-5.2 completed the replication workflow across all eight nodes in four of five trials, reaching seven nodes in the fifth.
However, the experiment substantially simplified several steps. The emulated Android nodes exposed an unauthenticated Android Debug Bridge (ADB) interface, so no exploitation was required. The experimental framework, not the agent, spawned each successor copy once a node passed correctness checks. The agent's reasoning ran on separate inference infrastructure, because the test nodes could not host a model of GLM-5.2's size. The payload the agent deployed on each node was not itself but a smaller model, Llama 3.1-8B. Anthropic's reasoning is that if GLM-5.2 can autonomously perform these steps, a future model small enough to run on the target hardware but with GLM-5.2's capability could replicate itself. These results establish a capability floor for autonomous replication, and even with limitations, the broader implications are substantial.
Anthropic's Forecast: Why It Deserves Scrutiny
Anthropic's headline claim is that small models may be only months away from being capable enough to self-replicate across commercial IoT devices. GLM-5.2 is treated as the capability floor required for replication. Plotting notable open-weight models of 30 billion parameters or fewer against the Epoch Capabilities Index (ECI), a general measure of model capability, Anthropic fits a trend of roughly 13.7 ECI points per year. Extending that trend, a model small enough to run on 4GB devices reaches GLM-5.2's capability level by the end of 2026. Anthropic concludes that "small weight-based replicators may be possible soon after."
Anthropic itself notes the limits of this estimate. ECI is a proxy for general capability, not a direct measure of replication ability. Models of 30 billion parameters and under can already complete many individual replication steps but make mistakes composing them end to end. When Anthropic gave Qwen3.5-9B access to a frontier model (Claude Opus 4.8) as an advisor during the inference-setup task, the advice produced no measurable improvement. The small model generally did not know what to ask, and did not benefit from unsolicited guidance either. The reasonable reading is that the forecast is a trend-based projection with wide uncertainty. It is a projection from a credible research team, grounded in measurable experiments, and it points in a direction defenders should take seriously.
Autonomous Replication, Adaptation, and the Importance of Persistence
To understand why this research is important, consider how a self-replicating AI system might operate. A conventional worm might follow a programmed process to identify a vulnerable target, gain access, transfer executable code, and repeat. A sufficiently capable AI-driven system could theoretically expand this process by evaluating unfamiliar configurations, identifying compatible software, selecting alternative deployment approaches, and using information gained during earlier attempts.
Anthropic's research decomposes replication into four practical stages: discovering targets, obtaining execution capability, establishing AI inference, and continuing propagation. The adaptation component may prove significant.
In one experiment, successive instances of Qwen3.5-9B inherited notes describing earlier attempts, failures, and strategies. Across six generations and 20 independent lineages, initial inference-setup success was 45%; subsequent generations achieved success rates ranging from approximately 61% to 84%. It demonstrated a form of accumulated operational knowledge. It showed that later instances could benefit from information retained by their predecessors. Anthropic describes this as a kind of Lamarckian evolution, in which parent instances pass on prompts, code, and notes to their successors.
Traditional persistence refers to malware surviving reboots, restoring access, or maintaining a foothold within a compromised environment. Adaptive persistence adds another potential dimension by preserving useful information and strategies across successive attempts or agent instances. An AI agent that fails to configure software on a particular device might document the problem and pass that information forward. Another instance could avoid repeating the same mistake. Over time, that capability could make propagation more reliable across diverse systems. Whether such behavior develops into persistent, self-sustaining malware remains uncertain. Nevertheless, inherited operational knowledge presents a credible additional concern for defenders.
How Fast, and How Big?
Anthropic also estimated what a near-term replicator of this kind would look like in practice, and the numbers are worth watching. Running on smart-TV-class hardware, a single inference-setup task consumes roughly 28 to 51 hours for Qwen3.5-9B and 71 to 150 hours for the unoptimized Gemma4-26B-A4B configuration. Replication would therefore proceed on a timescale of one to two days per hop, with the population doubling over days rather than minutes.
If a fleet the size of the Kimwolf botnet, which is over two million devices, were fully converted to running Gemma4-26B-A4B in groups of 15, Anthropic estimates roughly 133,000 model replicas generating about 80,000 tokens per second in aggregate. For comparison, a single NVIDIA H100 GPU can serve the same model at roughly 1,950 tokens per second, so the entire botnet would represent at most about 40 H100s' worth of inference. Anthropic's analogy is a slime mold or a swarm of fruit flies: slow, distributed, and difficult to eradicate.
These are rough estimates anchored to one hardware class and one historical botnet. A more sophisticated replicator would be opportunistic about which devices receive a model payload and which serve as footholds. But the baseline picture is the near-term concern, and that is persistence and reach.
From a Laptop or Smartphone to the Rest of the Network
How might autonomous AI-enabled malware affect an ordinary household or business? Consider a hypothetical scenario involving a compromised laptop connected to a home network. The laptop may share network connectivity with smartphones, smart televisions, printers, cameras, security equipment, and automation hubs. If malicious software gains execution capability on the laptop, it may be able to identify other reachable network devices or services. Depending on network configuration, it might encounter outdated firmware, exposed administrative interfaces, poorly protected services, or shared credentials. An AI-enabled malicious agent could theoretically evaluate these conditions and select among possible paths for further access.
Sharing a Wi-Fi network does not automatically grant one device access to control another. Lateral movement requires a viable access mechanism, such as an exploitable vulnerability, credentials, an insecure management interface, or an overly permissive trust relationship. The router plays an important role because they connect local devices to external networks and frequently controls firewall policies, network address translation, wireless security, and local network segmentation. A properly configured router does not need to be compromised for infected devices to communicate through it. However, compromising its administrative functions could offer additional opportunities.
Other pathways may involve cloud services. For example, a security camera and smart lock might both depend on an automation platform accessible through a mobile application. Compromise of that platform account, or a connected service with excessive permissions, could potentially expose several connected systems. This is not the same as malware automatically traveling from a camera into a door lock. It demonstrates how shared identities, service integrations, and weak authorization boundaries can create relationships between otherwise distinct devices.
What can a compromised device or identity communicate with, authenticate to, and ultimately control?
Cameras, Microphones, Sensors, and Smart Locks: When Cybersecurity Becomes Physical Security
The consequences of device compromise are not limited to computer performance or information theft. Connected cameras may capture images of family members, employees, customers, or sensitive facilities. Microphones can collect conversations and other environmental audio. Smart locks, alarm systems, and connected access-control equipment can influence physical access and safety. If the necessary permissions are obtained or an affected device is exploited, these capabilities could create privacy and security risks.
For example, unauthorized camera access could expose live, private, or recorded video. Compromised audio equipment might reveal sensitive conversations. Unauthorized privileged access to a smart-lock management system could potentially issue lock or unlock commands. These outcomes are conditional on the architecture and security controls involved.
On a deeper dive, in January 2024, researchers at the Massachusetts Institute of Technology's Computer Science and Artificial Intelligence Laboratory (MIT CSAIL) published findings showing how ambient-light sensors, when combined with screen information and computational techniques, could reveal aspects of user interactions.
The laboratory research examined a privacy exposure involving sensors not normally considered cameras. It did not demonstrate widespread malicious surveillance through commercial devices, but the lesson is that sophisticated analysis can sometimes extract sensitive information from ordinary sensors. For households and enterprises, the appropriate response is to understand the information each device collects, which applications can access it, how permissions are enforced, and what would happen if those controls failed.
Privacy engineering, information security, and physical security are interconnected disciplines.
More capable edge systems, including AI-enabled cameras, video-conferencing units, and production servers, may provide computational resources attractive to future autonomous malware. Anthropic specifically flags accelerator-equipped smart cameras and conferencing hardware as plausible targets, while noting that most such devices communicate through outbound cloud tunnels, not exposed inbound ports, and that no mass compromise of this device class has yet occurred.
The stakes become more significant when examining operational technology (OT). Industrial environments use programmable logic controllers (PLCs), human-machine interfaces (HMIs), supervisory control and data acquisition (SCADA) systems, sensors, actuators, and other equipment to monitor and control physical processes. A PLC might receive signals from temperature, pressure, motion, or proximity sensors and respond by controlling motors, pumps, valves, or electrical relays. Unlike a compromised business application, unauthorized industrial control actions can potentially affect machinery, production continuity, equipment integrity, and, most importantly, personnel safety.
Many PLCs are designed for deterministic control tasks and cannot realistically host a contemporary large language model. A credible AI-enabled attack scenario is more likely to involve a compromised engineering workstation, industrial gateway, HMI, or another capable intermediary. If that intermediary has legitimate connectivity and operational privileges, malicious instructions could potentially reach downstream control systems and influence physical equipment.
This is why industrial cybersecurity requires more than conventional endpoint protection. Network isolation, engineering change controls, tightly governed remote access, independent safety systems, and verified recovery procedures become essential. For electric utilities and other critical-infrastructure operators, these principles are already codified in frameworks such as NERC CIP and the IEC 62443 series; the question this research raises is whether those controls are implemented and verified.
NIST's Guide to Operational Technology Security, SP 800-82 Revision 3, addresses these considerations through an OT-specific approach that accounts for operational reliability and physical safety. On September 21, 2026, NIST released the initial public draft of Revision 4, which restructures the guide around the Cybersecurity Framework 2.0, expands guidance on asset management and network monitoring, and introduces zero-trust architecture guidance for OT environments.
Autonomous AI Capabilities: Researchers Are Taking It Seriously
Anthropic's findings become even more meaningful when considered alongside independent research examining improvements in AI autonomy. METR, an organization specializing in AI capability evaluations, has studied how the length and complexity of tasks completed autonomously by AI agents have changed over time. Its original research reported that the human-equivalent duration of tasks AI agents could complete with 50% reliability approximately doubled every seven months across the 2019–2025 evaluation period.
In January 2026, METR expanded its evaluation suite from 170 to 228 tasks. Its updated methodology estimated a faster doubling period of approximately 131 days for models released since 2023, although estimates remain sensitive to task selection and evaluation methodology. These figures do not measure how quickly an AI worm can spread, nor do they predict when autonomous malware will become operational. Instead, they provide evidence that AI agents have been improving at independently performing longer, multi-step technical work. That is important because autonomous replication depends on successfully completing interconnected tasks.
Related research at Carnegie Mellon University's CyLab Security and Privacy Institute has emphasized the challenges of securing heterogeneous IoT environments through trusted device identities, autonomous network defense, and accountability for data access and collection. CyLab's work offers a perspective that as connected environments become more complicated and autonomous threats become more capable, defensive systems may also need greater automation and stronger trust mechanisms.
In January 2026, security research organization Synthient reported that the Kimwolf botnet had compromised more than two million devices, primarily off-brand Android TV boxes, by scanning for exposed and unauthenticated Android Debug Bridge (ADB) services through residential proxy networks. Kimwolf was not the autonomous AI system described by Anthropic, but its significance is twofold. First, a large population of insecure connected devices can already support substantial malicious infrastructure. Second, the access mechanism Kimwolf exploited in the wild, an unauthenticated ADB port, is the same one Anthropic left open in its test network. The "trivial" initial-access step in Anthropic's evaluation is less artificial than it might first appear.
Together, these findings support a reasonable conclusion that autonomous AI capabilities are improving while extensive networks of inadequately secured connected hardware already exist. The precise point at which these developments could converge remains uncertain.
Cybersecurity Safeguards: What Should We Be Doing Now?
The encouraging part of Anthropic's research is that many of the most important protective measures are already familiar to cybersecurity professionals. The challenge is implementing them consistently across an expanding population of connected devices. For households, small businesses, and enterprise environments, six priorities deserve particular attention.
Know what is connected.
Maintain visibility into connected computers, mobile devices, routers, televisions, cameras, printers, automation hubs, and other network equipment.
Identify manufacturers, device models, firmware versions, management interfaces, and support lifecycles. Remove obsolete or unnecessary devices, especially those no longer receiving security updates.
For enterprises, include unmanaged IoT, operational technology, digital media systems, and vendor-maintained equipment within asset management and risk assessment processes.
Update, configure, and harden devices.
Install supported firmware and operating system security updates. Enable automatic updates where appropriate and verify that critical systems receive patches through established change-management processes.
Replace default passwords, restrict administrative accounts, disable unnecessary services, and eliminate exposed development or debugging interfaces such as ADB.
Where not required, disable remote administration, insecure discovery features, and unnecessary inbound network access.
Select devices from manufacturers that provide documented vulnerability disclosure processes, reliable security updates, and defined support periods. Raising the baseline for all devices shrinks the target pool for scripted and AI-driven botnets alike.
Segment networks and limit unnecessary communication.
Separate higher-risk IoT equipment from personal computers, sensitive business applications, and critical infrastructure.
For a household, a properly isolated guest or dedicated IoT network can provide a useful security boundary.
Enterprises should consider VLANs, firewalls, access-control policies, and restricted inter-segment communications.
Industrial systems require additional protection through appropriately designed security zones, controlled conduits, and separation of business IT from safety-critical OT environments.
NIST's SP 1800-15 provides practical guidance using Manufacturer Usage Description (MUD), a mechanism for restricting IoT devices to the network communications required for their intended functions.
Strengthen identity and access management.
Apply strong authentication, multi-factor authentication where supported, least privilege, and role-based access controls.
These requirements should extend beyond human users to service accounts, device identities, API credentials, automation platforms, and AI agents.
Review permissions granted to applications that manage cameras, smart locks, alarm systems, and other connected equipment.
An AI-enabled threat may become substantially more consequential when it gains access to an identity with excessive privileges.
Monitor network behavior and prepare for containment.
Anthropic identifies several potentially observable characteristics of weight-based replicators (newly infected devices receiving multi-gigabyte model payloads, sustained device-to-device traffic between machines that would not ordinarily communicate at that volume, and the scanning and outbound connection patterns of propagation itself). Anthropic notes these signatures should be visible in ISP flow-based monitoring, and that a replicator faces an inherent trade-off between speed and stealth. Meaning the faster it spreads, the more traffic it generates.
Security teams should baseline normal communications and investigate unexplained deviations.
However, not every large download or unusual traffic pattern indicates malicious AI. Effective detection requires context, behavioral analysis, and appropriate validation.
Incident response plans should account for devices that cannot run traditional endpoint detection software. Honeypots can help defenders capture a replicator's harness and payload for analysis once activity is suspected.
In environments where immediate isolation might disrupt physical operations, containment procedures should be coordinated with engineering and safety personnel.
Build resilience into both digital and physical operations.
Maintain verified backups, supported recovery images, and tested restoration procedures.
For physical access systems, industrial equipment, and other cyber-physical environments, evaluate whether independent safety mechanisms and manual operating procedures remain available during cybersecurity incidents.
A network disruption should not automatically become an uncontrolled physical safety event.
These safeguards do not eliminate every potential attack, but they reduce available entry points, restrict propagation, and improve the likelihood of successful detection and containment.
Can Autonomous AI Malware Be Contained?
Anthropic examines containment through a concept borrowed from epidemiology: the reproduction number, or R₀. Applied to cybersecurity, this represents how many additional devices a compromised device can infect, on average. The objective is to drive the effective reproduction rate below one, making sustained propagation difficult. From an enterprise security perspective, the strategy is to contain the compromise, restrict lateral movement, and prevent isolated infections from becoming widespread incidents.
Research provides some useful perspective. A 2019 study by Çetin and colleagues, published at the Network and Distributed System Security Symposium, examined responses to Mirai infections across a mid-sized European internet service provider (ISP). Researchers found that isolating infected customers within a restricted quarantine network, commonly called a "walled garden," and notifying them resulted in 92% remediation within 14 days. By comparison, natural remediation without notification ranged from 58% to 74% across the control group and two reference networks. Email notifications alone produced no observable improvement.
However, two important distinctions deserve consideration. First, the research examined conventional IoT malware, not autonomous AI replicators. Second, Mirai was largely non-persistent, meaning a simple device restart could eliminate the infection. This contributed to relatively high natural remediation rates. An autonomous replicator capable of maintaining persistence, adapting to defensive measures, or reinfecting devices within minutes could present a considerably different challenge. In such scenarios, relying on natural remediation would be an inadequate security strategy. Nevertheless, the study demonstrates how coordinated network-level intervention can significantly improve containment and remediation outcomes.
Looking ahead, defending against AI-enabled replication may require greater coordination among device manufacturers, ISPs, cloud service providers, enterprise security teams, and government authorities. Unlike traditional malware campaigns dependent on centralized command-and-control infrastructure, independently operating AI replicators could make coordinated takedowns substantially more difficult. Anthropic acknowledges that completely eliminating such threats could require intervention across every ISP hosting compromised devices.
For CIOs, CTOs, and CISOs, the implications extend beyond traditional malware defense. Early detection, network segmentation, restrictive access controls, and coordinated incident response become important when the threat can independently adapt, persist, and propagate across interconnected environments.
What Enterprise Leaders Should Take Away
Anthropic's research raises questions that extend beyond the deployment of AI inside an organization. Today, much of the enterprise AI governance conversation centers on authorized use cases, responsible adoption, model selection, data protection, productivity, and financial returns. Those considerations remain important; however, organizations must also evaluate how AI may change the capabilities of external adversaries, including their ability to investigate complex environments, automate technical workflows, and adapt during attempted compromises.
An effective response should be grounded in established cybersecurity and enterprise risk management principles. Leadership teams should evaluate their visibility into connected assets, the effectiveness of network segmentation, the quality of firmware and vulnerability management, and the security of machine identities and remote access. The effectiveness of these programs can be measured through operational indicators such as supported-device coverage, critical patch remediation times, unauthorized connectivity exposure, segmentation-policy compliance, and mean time to detect and contain anomalous activity. For industrial environments, those measures must also account for safety, availability, and recovery requirements.
Importantly, organizations should avoid making significant technology investments solely because a speculative threat has attracted attention. Security investments should reflect demonstrated exposure, business impact, threat likelihood, implementation costs, and measurable risk reduction.
Looking Ahead: The Importance of Digital Trust
Anthropic's research offers a glimpse of what autonomous AI capabilities may eventually enable. Small models are becoming more capable. Connected devices are becoming more computationally powerful. Software now controls functions once isolated to dedicated equipment. These developments bring substantial benefits, including more capable automation, smarter infrastructure, advanced diagnostics, accessible digital services, and improved operational efficiency. They also create new dependencies and security considerations.
For individuals, that means understanding and protecting the devices connected to our homes and personal lives. For organizations, it means integrating AI-related threat considerations into cybersecurity architecture, governance, infrastructure management, and enterprise risk oversight. For engineers and technology developers, it means designing hardware, software, networks, and control systems with security and resilience as fundamental requirements. As intelligence becomes more autonomous, the trustworthiness of the systems it can access becomes important.
The future of cybersecurity will also depend on how well we protect our digital and physical environments from autonomous capabilities operating beyond our control.
Research and References
Primary research: William Zhou and Daniel Freeman, Anthropic Frontier Red Team, AI models may be capable of self-replicating without access to a datacenter, October 12, 2026.
Additional research and guidance:
MIT CSAIL, Study: Smart devices' ambient light sensors pose imaging privacy risk, January 2024.
Carnegie Mellon University CyLab, Secure and Private IoT Initiative.
METR, Measuring AI Ability to Complete Long Software Tasks, March 2025; and Time Horizon 1.1, January 2026.
NIST, SP 800-82 Rev. 3: Guide to Operational Technology Security, September 2023.
NIST, SP 800-82 Rev. 4 (Initial Public Draft): Guide to Operational Technology Security, September 21, 2026; comment period through November 30, 2026.
NIST, SP 1800-15: Securing Small-Business and Home IoT Devices, May 2021.
Synthient, A Broken System Fueling Botnets, January 2026.
Orçun Çetin, Carlos Gañán, Lisette Altena, Takahiro Kasama, Daisuke Inoue, Kazuki Tamiya, Ying Tie, Katsunari Yoshioka, and Michel van Eeten, Cleaning Up the Internet of Evil Things: Real-World Evidence on ISP and Consumer Efforts to Remove Mirai), NDSS Symposium, 2019.
UK National Cyber Security Centre, Smart devices: new law helps citizens to choose secure products.
U.S. Federal Trade Commission, Securing Your Internet-Connected Devices at Home.
Editorial disclosure: This article represents independent research commentary by DEW Diligence, LLC. DEW Diligence is not affiliated with or endorsed by Anthropic, MIT, Carnegie Mellon University, METR, NIST, Synthient, or the other organizations referenced. Model names, hardware configurations, and performance figures are reported as published by Anthropic and have not been independently reproduced by DEW Diligence. Descriptions of potential future threats are identified as hypothetical and should not be interpreted as documented incidents. "Adaptive persistence" is a DEW Diligence framing and is not terminology used by Anthropic. Anthropic's research is on "Artificial Intelligence" and "Autonomous Replication and Adaptation (ARA)." "Superintelligence" is terminology used by DEW Diligence.
Want more on this topic? Contact us for a consultation, and review our services for fractional & virtual CIO, CISO, and CTO executive consulting.
David E. Williams, CISM, GSTRT, GSLC
Founding Principal, DEW Diligence, LLC
© 2026 DEW Diligence, LLC.
Comments