The question in our title sounds like a routine IT-policy problem. Anthropic just gave it a far stranger answer. In a research post, researchers William Zhou and Daniel Freeman wrote that small AI models may be months away from being capable enough to self-replicate on commercial IoT devices, and that such a replicator 'would be challenging to shut down, since there is no centralized point to shut off'. That last clause is the whole lesson, and it is one most operators are not set up to absorb.
Why the off-switch lesson reaches beyond AI worms
The Anthropic work is about a specific and frightening class of malware: a self-replicating model that spreads across cheap internet-connected devices and runs inference directly on the hardware it has captured. But the underlying pattern, no control to pull, shows up anywhere an autonomous tool runs somewhere you cannot see. The question in our title is not really about AI. It is about whether you can name, count, and kill every process that carries out work in your company without asking you first.
Anthropic's own framing makes the point. Their 'AI worm' carries its model weights as part of the payload, or re-downloads them from a public hub, then runs inference on the machines it has taken over, "effectively, the weights would become part of the malware payload." There is no headquarters to bomb. The replicator is the network.
What the experiment actually showed
Before we lean on the scary part, here is the factual part, because the facts themselves are already useful to a risk officer. Anthropic ran small open-weight models on real and simulated low-cost hardware. Their demo distributed the Gemma4-26B-A4B model across 15 off-brand Android TV boxes, each with 4GB of memory, running at about 0.39 tokens per second. A smaller Qwen3.5-9B model managed to set up inference on eight of nine hardware types they tested, according to the post. The company reported that a larger model, GLM-5.2, could complete the full replication cycle reliably, spreading across eight nodes in four of five test runs.
Two details matter for operators. The replication chain breaks down into four steps: finding targets, gaining execution, establishing inference, and propagating to a new host. Of those steps, the third, setting up inference on unfamiliar hardware, is the least studied and may be the most challenging for small models, because "these steps are serially dependent, and failures are difficult to recover from". And the whole thing is slow. In Anthropic's setup, a full inference setup could take between 71 and 150 hours for the 26B model, and 28 to 51 hours for the 9B model.
The part that should keep executives awake
The scary part is not the current numbers. It is the trend line. Anthropic estimated that the best small models improve at a rate of about 13.7 points on a capability index per year, and that by this trend a small model could reach the replication floor by the end of 2026, with weight-based replicators possible "soon after." The model also learns on the job: each failed attempt can leave prompts and notes behind so the next copy avoids the same dead end, a process the post describes as a kind of Lamarckian evolution.
There is a crucial caveat, and it is worth quoting because it protects the whole argument from hype. Independent researcher Xiaohan Fu, interviewed by Ars Technica, said the study "does not, on its own, establish that small models would autonomously choose to replicate." In the actual experiment, the replication was orchestrated by a larger model while the target devices ran a smaller one, and the target devices exposed an unauthenticated Android Debug Bridge, a development tool that should never be reachable in production.
The mirror it holds up to your company
This is where the article answers the title. Anthropic's insight is that some threats have no central point to shut off, so you must assume the off-switch will not work and design around that. Now point the same question at your own agentic tooling.
Ask three things. First, can you inventory them? If your company runs agentic coding tools, procurement automation, cloud provisioning agents, or support bots that can open new accounts, many of those decisions now happen on devices and cloud regions you may not be logging. The IoT lesson is that cheap, poorly defended compute quietly multiplies outside your radar.
Second, where is the kill switch, and does it do anything? Anthropic noted that their replicator need not need root privileges, because inference can run entirely in userspace and devices can be re-infected on restart. The analogy is direct: if an autonomous tool, or the malware that rides along with one, lives on a device outside your management plane, your big red button may only restart the problem. Third, how fast would you even know? A replicator the size of Anthropic's demo spreads over days, not minutes. By the time an incident response team gets called, the process has already branched to machines that do not appear on your dashboard.
What control actually looks like now
The honest answer to the title is that your control is probably much less than your architecture assumes, and that is fixable if you treat it that way. You cannot assume a single off-switch exists, so you have to distribute the defense: map every place autonomous agents can run, including third-party SaaS and any device that connects to your network, constrain what those agents are allowed to reach with the tightest network and credential boundaries you can manage, and assume that detection will lag by days.
Anthropic's research is a warning about malware with no central host. Your agentic tools are the same shape from the inside, only authorized. The goal is not to stop using them. It is to build an organization that would still survive the day you could not find, let alone disable, one of them running somewhere it was never supposed to be.
Sources
- AI models may be capable of self-replicating without access to a datacenter (Anthropic, 2026-10-14)
- Anthropic warns small AI models could self-replicate through IoT devices (Ars Technica, 2026-10-14)