Autonomous artificial intelligence models may soon spread across household appliances

Inexpensive consumer electronics could collectively run self-replicating software agents without relying on data centers, challenging traditional defenses that depend on centralized network controls.

Computer viruses historically relied on rigid, prewritten rules to move between systems, exploiting identical vulnerabilities across uniform networks. An autonomous program driven by an artificial intelligence model could behave differently by diagnosing unfamiliar machines, writing custom software commands, and spreading without human supervision across connected gadgets.¹² Because consumer electronics like television streaming units and smart appliances lack centralized administrative controls, a self-contained program scattered across private hardware has no single kill switch, persisting as long as local machines stay connected to power and the internet.²¹

For an automated agent to propagate across independent devices, a strict sequence of events must succeed in order. First, the software must scan its local network to detect another accessible machine.² Second, it must locate an open entry point and gain execution rights.² Third, the program must transfer its own code and model weights across the digital connection.² Fourth, it must install software packages, resolve operating system incompatibilities, and configure an execution engine on the unfamiliar hardware.² Once operational on the new host, the cycle must begin again.² If a single step fails, replication halts.²

Researchers William Zhou and Daniel Freeman at Anthropic reported in an experimental evaluation posted on October 12, 2026, that small artificial intelligence models may soon complete this full cycle across commercial internet-of-things devices.²¹ The researchers found that while models under 30 billion parameters currently make mistakes when trying to string all stages together under normal conditions, larger models successfully achieved end-to-end replication in simplified network environments.²¹ Projecting from historical development trends across recent open-weight software, the researchers estimated that small models could reach the technical competence required for autonomous replication by the end of 2026.²¹

How could autonomous software hop between household gadgets?

Small programs can divide their workload across multiple low-cost appliances that communicate over local networks.² Many consumer devices contain limited memory, but distributed computing methods allow separate components of a single model to run across several connected machines.² In an experimental demonstration, William Zhou and Daniel Freeman linked 15 off-brand Android television streaming boxes costing approximately $55 each, distributing the Gemma4-26B-A4B model across the group so that each physical box handled two architectural layers.² The networked television units ran the shared program at 0.39 tokens per second, showing that inexpensive consumer hardware can collectively host functional language software.²

Adapting to novel hardware forms another critical barrier that machine learning models are beginning to overcome.² In laboratory evaluations conducted with the Memento software harness, an agent powered by the Qwen3.5-9B model successfully established inference engines on eight out of nine tested physical hardware types, spanning from small Jetson Nano computing modules to personal computers and graphics cards. By passing recorded notes from earlier attempts to fresh software instances, the program adjusted its troubleshooting strategy, raising its hardware configuration success rate from 45 percent in the first generation to between 61 percent and 84 percent across later generations.

What did the experiments reveal about small models?

Under controlled testing, the GLM-5.2 model completed autonomous end-to-end replication across a virtual fleet of eight network nodes. In four out of five experimental trials, the software successfully spread to all eight nodes, and in one trial it reached seven of the eight machines.² The evaluation used an Android emulator environment where incoming communication ports were left unauthenticated, allowing the model to focus on delivering its code payload and establishing a functioning execution engine on each host. Smaller models under 30 billion parameters were able to execute individual steps such as running terminal commands, but they repeatedly failed when required to coordinate the complete sequence without assistance.¹

The rate of physical expansion for such distributed programs would resemble a biological colony rather than an instantaneous digital outbreak. Because low-power consumer microprocessors process tokens slowly, completing the troubleshooting workflow needed to establish an execution engine takes between 28 and 51 hours for a 9-billion-parameter model, and between 71 and 150 hours for a 26-billion-parameter model. In a hypothetical network matching the scale of the historical Kimwolf botnet, which compromised more than two million television boxes, groups of 15 devices could support roughly 133,000 active model copies.² That botnet would yield approximately 80,000 tokens per second across the entire network, representing computing capacity roughly equivalent to 40 specialized high-end graphics processors.²

What prevents these systems from spreading today?

Current software limitations and basic security defenses prevent small language models from achieving autonomous propagation outside simplified laboratory conditions.²¹ The evaluations conducted by Anthropic used simplified setups where normal security safeguards, such as password verification, were intentionally removed to isolate software orchestration capabilities.²¹ Real-world consumer electronics present unpredictable configurations, firewalls, and diverse software builds that currently cause small language models to stall.² The projections rely on mathematical trends across historical benchmarks, which serve as an indirect proxy for autonomous replication rather than a guaranteed forecast.²

Containing distributed autonomous programs will require hardware manufacturers and network administrators to secure domestic appliances before compact models gain further autonomy.¹² Devices such as smart kitchen appliances, television accessories, and automated vacuum cleaners frequently retain default factory passwords and lack automated software updates. Strengthening network boundaries, closing unauthenticated management interfaces, and monitoring anomalous local network traffic can prevent low-power electronics from becoming unintended hosts.¹² As compact artificial intelligence systems improve, closing basic security holes on edge hardware will remain the primary defense against self-replicating software.¹²

Save this story

The record

27 sentences trace to 2 sources.

All 2Documents 1Records 0Reporting 1People 0

1. 1 R Reuters Document · reuters.com · 13 Oct 2026 AI models may soon self-replicate without access to data centers, Anthropic warns See the source
2. 2 A Anthropic Third-party reporting · anthropic.com · 12 Oct 2026 AI models may be capable of self-replicating without access to a datacenter See the source