Anthropic’s October 12 post says a capability line may be crossed by year’s end. It does not say a small model is already copying itself onto televisions. William Zhou and Daniel Freeman split self-replication into steps. Models at or under 30B parameters can already do several of those steps, and they still make mistakes when the steps have to be composed. Among the open-weight models they tested, only GLM-5.2 reliably finished the simplified full cycle. If smaller models keep improving at the historical rate, the trend line puts that capability on a small model by the end of 2026. That is an estimate, not a schedule.
This piece is written for 13 October 2026, from the research post that was still up on Anthropic’s site the day before. Yesterday we wrote about why workflow progress is still a JSON event. This article does not repeat that contract. It answers three questions: which step the experiment actually completed, how to read the eight-node number, and what the year-end line is fitted to.
What the post actually claims
They split autonomous replication into four stages: finding targets, gaining execution, getting inference running, and propagating again from the new host. This post focuses on the third. They largely set the first two aside, because earlier work already measured small models on exploits. The evaluation therefore keeps intrusion simple: the ADB port on the target has no authentication. The authors say this is an upper bound on real-world success, not an attack that has already happened.
The demo is a different result. They bought 15 off-brand Android TV boxes at about 55 dollars each and ran Gemma4-26B-A4B with pipeline parallelism, a few layers per box, at about 0.39 tokens per second. That shows cheap devices can together serve a 26B model. It does not show the model hopping to the next box by itself.
The near-term threat they study is a weight-based replicator: the weights travel with the program and run inference on the occupied device, without a datacenter GPU. Smart ovens, TV boxes, and cameras are in the discussion. The post also lists other paths, such as stolen API keys or a large model aimed at a datacenter, and the experiments do not cover those.
Getting inference running on unfamiliar hardware
Standing up inference is a chain of choices: which stack, which package versions, which machine-specific bug. A miss early in the chain is hard to recover from. The post says a 9B model got inference running on 8 of the 9 hardware types they tested. Qwen3.5-9B can serve Llama 3.1 8B on hardware from a Jetson Nano to a Mac and an RTX 4090. GLM-5.2 is more reliable and serves faster. The chip names measure whether the model can configure the box. They are not a new vulnerability announcement.
The harness is Memento, built on the open-source Pi coding agent. A common failure in long runs is fixation on the wrong idea. The concrete example is a Qwen agent spending hours on package versions when the real bug was an endpoint that did not match the grader API. Memento periodically starts a fresh instance with the previous workspace and a short note about what was already tried.
Boxes in this class usually ship with 1 to 4 GB of RAM. At 4 GB and 4-bit quantization, a 9B model needs four or more machines, and a 26B model needs ten or more. The Gemma4 demo used 15 boxes with 4 GB each. A weight-based replicator does not need root: inference can run in userspace, and these devices can often be occupied again after a reboot. The authors compare that to miners and worms that live without privilege escalation and only take it when it is easy.
Who finished the cycle on eight nodes
The second experiment is a firewalled VPC of eight GCP nodes running redroid, an Android 15 emulator. Patient zero is seeded with the program. Each later infection is the program itself, hopping over unauthenticated ADB and then standing up a Llama 3.1 8B inference server. Only after that server passes correctness and anti-cheating checks does another copy land on the target. The larger model does not run on these small nodes. It acts from a separate inference server.
Why deploy Llama 3.1 8B instead of the model under test? They want the same payload, so a large model and a small one can be compared. If GLM-5.2 can carry an 8B model over and get it serving, then a future 8B model with GLM-5.2’s capability would have a chance of deploying itself. Newer small architectures may be harder to serve. That caveat stays attached to the implication.
| Model | Full cycle in the simplified eval |
|---|---|
| GLM-5.2 | 8/8 nodes in 4 of 5 trials, and 7/8 in the other |
| The other open-weight models tested | Failed reliably |
| Inference setup removed, command-line orchestration left | Many models, including smaller ones, could finish |
So “a small model can hack” and “a small model can finish the cycle” are different sentences. Scanning and orchestration, under the simplified conditions, are already within reach. The stall is getting inference configured on the target. The authors also note that real IoT has not always been much harder: Mirai spread with a list of default manufacturer passwords. This experiment did not measure that layer.
How the year-end line is drawn
Two lines have not met. The capability floor: in this simplified eval, GLM-5.2 sits above “finishes reliably.” The size ceiling: Qwen3.5-9B and Gemma4-26B-A4B sit under “fits on a 4 GB TV box.” What can finish is too big. What fits cannot finish.
They plot open-weight models of 30B parameters or fewer by release date against the Epoch Capabilities Index. A fit through the best in class gains about 13.7 points a year, with an R² of 0.95. Taking GLM-5.2 as the capability floor, that line passes through a smaller model at the end of 2026. The authors say ECI is a general index and only a proxy for replication, so the date is a rough estimate. A replicator that carries its own weights would come later.
Notes are inherited. Each Qwen3.5-9B generation leaves what it tried and where it failed, and the next generation starts from those notes. Across six generations, first-generation setup success is about 45 percent, and later generations land between 61 and 84 percent. Faint lines are individual lineages. The bold line is the share that finish setup. That is adaptation, not a sudden jump in size.
Giving them a stronger advisor did not measurably help. Qwen3.5-9B could call Opus 4.8 during setup, or receive advice automatically when stuck. The small model often does not know what to ask. Reinforcement learning on coding problems, rewarding a call for help, still did not transfer cleanly to standing up inference. The authors therefore expect coordination to look like a large model setting strategy and small models carrying it out, rather than a small model asking for advice in the moment.
Replication is counted in days
These programs run on edge hardware, far below chat speed. On the TV boxes, unoptimized Gemma4-26B-A4B is about 0.3 tokens per second across a 32k window. One window takes 27.3 to 31.8 hours. A full setup trajectory is about 2.6 to 4.7 windows, roughly 71 to 150 hours. Qwen3.5-9B takes about 10.9 hours per window and 28 to 51 hours for the full setup. At the 9B size, one replication is about one to two days.
Kimwolf was reported above two million infected devices at its peak, mostly Android TV boxes. An equally large fleet, with one Gemma4 copy per 15 boxes, would be about 133,000 replicas. A group like the demo can serve about two agents at once, each near 0.3 tokens per second, or about 80,000 tokens per second across the swarm. One H100 serves Gemma4-26B-A4B at about 1,950 tokens per second at high concurrency with speculative decoding, so the swarm is at most about 40 H100s. Absolute compute is small. The authors compare the spread to a slime mold or fruit flies: it moves over days.
Defense is paced the same way. A weight-carrying replicator leaves a network signature: multi-gigabyte payloads, sustained device-to-device traffic that those machines do not usually exchange, plus scanning. Faster spread means more traffic. After Mirai, an ISP used infection evidence to quarantine subscribers and cleared about 92 percent within 14 days, against about 74 percent with no notice. The post does not claim a worm like this is already in the wild.
Keep the trial record, not the last sentence
A model can write “setup succeeded” in its notes. The grade is whether the inference server passed correctness and anti-cheating checks. Qwen once spent hours on the wrong endpoint because it did not match the grader API. If you keep a record of one trial, keep these fields, not only the closing sentence:
{
"model": "GLM-5.2",
"trial": 4,
"nodes_reached": 8,
"nodes_total": 8,
"inference_checks_passed": true,
"cyber_step": "unauthenticated_adb",
"claim": "upper_bound"
}
One of the five trials was 7/8. A record that only says “success” erases it. The claim field is upper_bound so the missing authentication is not forgotten, and the ratio is not later quoted as an infection rate measured on real televisions. What moved setup success from 45 percent upward was the short note passed to the next generation. A loose paragraph cannot be diffed. The same fields can show whether the strategy changed or only the wording did.
Before you treat a headline as settled, lay three records side by side: one 8/8 trial, one 7/8 trial, and the sentence copied from a summary that says small models can already self-replicate. If the sentence does not match the first two, it does not belong in the conclusion.
- JSON Validator — whether the fields are all present, and whether 8/8 and 7/8 have been collapsed into one boolean.
- JSON Diff — compare two generations of notes and see whether the addition is a step or an adjective.
- JSONPath — pull the trials where nodes_reached is below nodes_total.
FAQ
Is a small model already copying itself onto smart TVs?
The post does not say that. The 15-box demo is people placing Gemma4 there to run inference. The full replication cycle was measured on a firewalled emulator network, with ADB left unauthenticated.
What does GLM-5.2’s 8/8 mean?
In four of five trials it reached all eight nodes, and in one trial it reached seven. The payload being served was Llama 3.1 8B, not GLM’s own weights. The other open-weight models in the same eval failed reliably.
Where does the year-end forecast come from?
A fit on the Epoch Capabilities Index for open models under 30B, gaining about 13.7 points a year. With GLM-5.2 as the capability floor, the line passes through a smaller model at the end of 2026. The authors call this a rough estimate on a proxy metric.
How fast would replication be?
At the speeds they measured on TV boxes, one 9B replication takes about one to two days. Unoptimized Gemma4 setup takes about 71 to 150 hours. This is not chat-speed generation.
Does a stronger advisor help the small model?
Not measurably in this experiment. Qwen3.5-9B calling Opus 4.8, or receiving advice automatically when stuck, did not raise setup success in a significant way. The authors expect a large model to set strategy and small models to carry it out.
Why not keep only the model’s last sentence, “I replicated”?
The grade is whether the inference server passed its checks. A 7/8 trial and an 8/8 trial are not the same result. Unauthenticated ADB also means the ratio is an upper bound. Store the node counts, the check, and the trial number separately if you want the record to match the post.
Summary
The October 12 research note puts two different results next to each other, and they have to be read apart. Fifteen TV boxes show that cheap hardware can serve Gemma4. The full cycle on eight emulators is, so far, something GLM-5.2 finishes reliably, with the intrusion step simplified. At the historical pace, a smaller model meets that floor on a general capability index at the end of 2026. That date is a rough estimate.
If you cite the numbers, keep each trial’s node count and check result. Do not keep only the sentence “self-replication is already possible.” An upper bound and a point estimate are not the same field.