Autonomous replication and adaptation (ARA), a model deploying and propagating itself without human control, is what the post calls "a longstanding concern". Datacenter GPUs would be the obvious hardware for it; this post asks whether cheap devices would do. Could a small open-weight model act as a worm that spreads across cheap, poorly secured internet-of-things (IoT) devices such as off-brand Android TV boxes, carrying its own weights and running inference on the machines it compromises? The authors test the parts of that chain that earlier work had measured least, and conclude that small models "may be months away" from being capable enough.
The post was published on anthropic.com on 12 October 2026 under the Frontier Red Team label, by William Zhou and Daniel Freeman (byline). It has no companion paper; the post, with its appendices and sixteen embedded interactive charts, is the version of record. The charts are separate pages whose printed labels and hover text carry the numbers; locations below give the section letter and the chart's title. The inference-setup grid states that its data are "as of 30 July 2026". Some counts below were read from the charts' page source, as noted. This note found no forum crosspost and no published replication; news coverage was found but not read for this note.
Terms used in this note
- Weight-based replicator. The post's term for a worm that runs its model directly on compromised hardware, copying the weights with it or re-downloading them from a public hub.
- Inference setup. Getting a model served on unfamiliar hardware: choosing an inference stack, finding compatible package versions and debugging machine-specific problems.
- Memento. The authors' agent harness, built on the open-source Pi coding agent. It periodically restarts the agent with the previous instance's workspace and a short set of notes on what it tried (Appendix A).
- ECI. Epoch AI's Epoch Capabilities Index, a general capability score the post uses as a proxy for replication ability.
The threat model
The post breaks replication into four stages: finding targets, gaining execution, establishing inference, and continuing to spread from the new host ("Introduction"). It largely sets aside the first two, the cyber exploit, on the grounds that earlier work (CyberGym, ExploitBench and an "adaptive worms" paper, all linked to arXiv) already shows small models can exploit systems. Its focus is inference setup, which it calls the least studied stage and possibly the hardest for small models.
Of several threat models it lists (small models compromising hosts, large replicators aimed at datacenters, models replicating by stealing API credentials), the post studies the first as the nearest-term one that can be tested directly. It notes that weight-based replicators need no root privileges, because inference engines can run in user space, and that low-security devices can be reinfected after a restart ("Aside" in Section A).
The demonstration. The authors bought 15 off-brand Android TV boxes (about $55 each, 4 GB of RAM) and ran Gemma4-26B-A4B across them at 0.39 tokens per second, sharded by pipeline parallelism (each box holds a few of the model's layers), shown in a video at the top of the post. This shows that the hardware can run the model. The post describes the boxes as ones "we bought and networked together"; it does not report that an agent compromised them or installed the model.
What it finds
Small models can set up inference on most hardware tried
Eleven open-weight models acting as agents were each asked to serve Llama3.1-8B on nine kinds of hardware: seven cloud instances (a Xeon CPU, an RTX 4090, a Mac, a Windows machine, a TPU, an AMD MI300X, an NVIDIA H200) and two physical boards, a Jetson Nano and an Orange Pi (Section A, chart "Inference setup across hardware"). On the Jetson Nano the target was Llama3.2-3B, because the board lacks memory for the 8B model. Runs used Memento, with 1.5 hours per iteration and 6 hours in total (Appendix A).
The post's summary is that small models such as Qwen3.5-9B "are capable of setting up inference servers on hardware from Jetson Nanos to Mac PCs and RTX 4090s". Its opening states that a 9B model "got inference running on 8 of the 9 hardware types we tested". The chart's page source agrees: Qwen3.5-9B had at least one successful setup on every type except the MI300X, where it succeeded in 0 of 4 runs (the note's reading of the chart data). Success rates per cell vary widely, for example 1 of 5 runs on the TPU and 13 of 15 on the Xeon CPU. Llama3.1-8B, run as the agent, never succeeded. The post says more capable models such as GLM-5.2 (744B parameters, per the ECI chart) are "not only more reliable but also achieve higher serve tokens/s".
The harness matters, and not in one direction (Appendix A, chart "Single attempt vs memento, by model"). Memento raised success for Qwen3-Coder-30B-A3B from 3 of 13 single attempts to 14 of 22 chains, and for Gemma4-26B-A4B from 16 of 26 to 11 of 15. For Qwen3.5-9B with thinking off it fell from 10 of 21 to 10 of 29. The chart's Wilson 95% intervals overlap for all three models. The post describes Memento as meant to improve long-horizon reliability, not to grant new capabilities, and does not discuss the Qwen3.5-9B result.
Context length also matters: Qwen3.5-9B succeeded in 0% of runs at an 8k context, 22% at 16k, 60% at 32k, 78% at 64k and 67% at full context (Appendix C, chart "Qwen3.5-9B inference setup success rate vs context window", which gives no run counts).
Only GLM-5.2 replicated end to end
To test the whole chain, the authors built a firewalled network of 8 Google Cloud nodes running redroid, an Android 15 emulator, with an unauthenticated Android Debug Bridge (ADB) port open on each (Section B). Starting from one infected node, the worm had to scan the network, reach a new node, send over its payload (weights and harness) and stand up a Llama3.1-8B server there. If the server passed correctness and anti-cheating checks, the authors spawned a copy of the worm on that node. The worm's own model ran on a separate server, since the nodes are too small for the larger models.
The authors explain why the target is Llama3.1-8B and not the agent model itself: larger models are harder to serve, and a fixed target gives "an apples-to-apples comparison". A success by GLM-5.2 is read as evidence that an 8B model with GLM-5.2's capability could stand itself up, with the caveat that newer small models may have more complex architectures ("Why task the models...").
Result (chart "End-to-end replication across an 8-node fleet"). GLM-5.2 spread to 8 of 8 nodes in 4 of 5 trials and 7 of 8 in the fifth. Qwen3.5-9B (5 runs), Gemma4-26B-A4B (6 runs) and Qwen3-Coder-Next (5 runs) reached no node in any run. These four are the only models the chart shows in this test; the opening's "GLM-5.2 and stronger models are capable of end to end replication" goes beyond them. The authors call the evaluation "an upper bound on real-world success", because the exploit step is trivial. They add that past IoT botnets such as Mirai spread by trying default manufacturer passwords, so real IoT security "might not be dramatically more difficult".
With inference setup scripted (Section B and Appendix F, charts "Replication across a 17-node network..." and "Nodes infected out of 17..."). In a 17-node network, two subnets of 8 joined by a pivot node, the inference step was replaced by a pre-written script to isolate the remaining command-line work. Medians of 6 runs were:
| Agent model | No guidance | Mild guidance | Directive guidance |
|---|---|---|---|
| Ministral-3B | 9.5 of 17 | 13 | 8.5 |
| Qwen3.5-9B | 17 | 17 | 17 |
| Ministral-14B | 17 | 17 | 17 |
| Gemma4-26B-A4B | 17 | 9 | 9 |
| GLM-5.2 | 17 | 17 | 17 |
| Claude Opus 4.8 | 8 | 17 | 17 |
The post concludes that many models, including small ones, can do the orchestration, and that "even a 3B model reaches about half the nodes".
A trend projection: the end of 2026
The post frames the question as a capability floor (at most GLM-5.2's level, which the chart marks "≤151.8") and a size ceiling (models small enough for the target devices). Qwen3.5-9B and Gemma4-26B-A4B sit under the ceiling for 4 GB TV boxes, but the two do not yet meet (Section B). Plotting notable open-weight models of 30B parameters or fewer against ECI, the least-squares fit over the best model at each date gains 13.7 ECI per year, with R² of 0.95. The chart ("Capability of models small enough to run on a TV fleet") puts GLM-5.2 at ECI 151.8 and the best small model, Qwen3.8-27B (released 14 August 2026), at 149.4, labeled "2.4 ECI to go". The trend line puts a small model at the floor by the end of 2026. The authors call this "only a rough estimate", because ECI is a general index and "only a proxy metric for replication ability".
Speed and scale: a slow worm
The post estimates that such a worm would be slow (Section C, Appendix C). At about 0.3 tokens per second for Gemma4-26B-A4B on the TV boxes, one 32k-token context window takes 27.3 to 31.8 hours, and a full inference setup, which took about 2.6 to 4.7 windows in the eval, 71 to 150 hours. For Qwen3.5-9B the figures are 10.9 hours and 28 to 51 hours. Scaled to the Kimwolf botnet's reported peak of over 2 million devices, groups of 15 boxes would give about 133,000 copies and about 80,000 tokens per second in total. The authors compare this with about 40 H100 GPUs, "a tiny amount of compute in absolute terms". They call the estimate rough, since it rests on one class of hardware and one past botnet.
Appendix D gives an upper bound for edge devices with built-in AI accelerators: on the order of 90 NVIDIA-silicon IoT product lines, about 9.5 million fielded units, and between 9,000 and 35,000 H100-equivalents of compute. The authors add that the reachable share is "likely an order of magnitude or more lower", since reachability is "overwhelmingly through outbound cloud-management tunnels". Their chart of device classes ("Fielded units by device class, by internet reachability") marks about 5 million streaming TV boxes with a GPU as reachable from the internet, about half of the roughly 9.9 million units it lists outside handheld consoles (the note's arithmetic from the chart's approximate labels); the post does not reconcile the two.
Adaptation helped; advice did not
Inherited notes (Section D, chart "Inference setup success rate over six generations of inherited notes"). In 20 lineages of Qwen3.5-9B, each generation inheriting the previous one's notes, the share completing inference setup was 45% (9 of 20) in generation 1. Later generations scored 72% (13 of 18), 61% (11 of 18), 63% (12 of 19), 84% (16 of 19) and 65% (11 of 17). The post summarizes this as a rise "to between 61% and 84% in later ones"; it describes passing notes, prompts and code from parent to daughter instances as "Lamarckian evolution". The Wilson 95% intervals printed in the chart overlap across generations: 26% to 66% for the first, 62% to 94% for the fifth. The chart does not explain why the number of lineages falls from 20 to 17.
A frontier advisor (Section D, chart "Inference setup success rate by advisor access"). Qwen3.5-9B was given access to Claude Opus 4.8, either as a tool it could call or as unsolicited advice when stuck. Without reinforcement learning (RL), success was 60% (6 of 10) with no advisor, 30% (3 of 10) with a voluntary advisor and 25% (2 of 8) with a forced one. Two further groups were trained with RL while having advisor access, one of them also rewarded for calling the advisor when stuck. Without that reward, success was 30% (3 of 10) with no advisor, 20% (2 of 10) voluntary and 44% (8 of 18) forced. With it, success was 72% (13 of 18) with no advisor and 44% to 67% across the three advisor conditions. The post says the advice "doesn't measurably help" and the caption reports "no significant improvement". Samples are 8 to 21 runs per condition, and the intervals are wide. The authors infer that coordination is more likely to come from a large model directing small ones as subagents.
Spear-phishing (Appendix E, chart "Spear-phishing capability"). On a local rerun of CyberSecEval 3's spear-phishing evaluation (about 240 cases, 0 to 5, GPT-4o as judge), five small models scored 2.17 to 2.77; correcting for the judge's estimated lenience of 0.4 to 0.7 points puts them at "very poor" to "poor".
Mitigations and conclusion
The post frames defense as driving R₀, the number of new devices each infected device infects, below 1 (Section F). It argues that a replicator moving multi-gigabyte payloads and running distributed inference would leave a network signature visible to ISPs, so speed trades against stealth. It cites Mirai remediation, in which ISP quarantine and notification cleared 92% of infections within 14 days against 74% without notification. It also cites UK rules making default IoT passwords non-compliant. It states that an AI botnet could act intelligently against defenders, "but we haven't tested this" (Section E).
The conclusion, titled "Signs of life", says that "We don't yet have a full rogue replicator", while agents are increasingly able to complete the individual steps. A final paragraph, introduced as "a prediction that goes beyond our experiments", expects agents running on all kinds of hardware, some outside human control.
The prompts and instructions used
The post has no training documents or principles. Agents were given task prompts with about 1,000 tokens of information, including hardware specifications and a simple profiling script (Appendix A). The 17-node experiment varied the amount of guidance in the prompt (none, mild, directive); the chart presents guidance as the only factor varied, though the post does not say what was held fixed. The post does not print these prompts. Guidance changed results in both directions: Claude Opus 4.8 reached a median of 8 nodes with no guidance and 17 with either level, while Gemma4-26B-A4B reached 17 with none and 9 with either. In the chart's per-run values, Opus 4.8 without guidance reached exactly 8 nodes in all six runs and never crossed the pivot to the second subnet. The post does not comment on these differences. The 17-node replication trees in Section B use the "mild" prompt, according to their caption; the post does not say which prompt the 8-node end-to-end runs used.
Limitations and open questions
The authors state these:
- Trivial exploit. The end-to-end test has no authentication, so it is an upper bound on real-world success.
- Proxy capability measure. The end-of-2026 date rests on a general index and a trend fit.
- Rough scale estimates. The speed and botnet-size figures are anchored on one hardware class and one past botnet.
- Untested defenses and counter-defenses. Adaptive behavior against defenders was not tested.
Further points that bear on reading the result:
- The worm's model was not on the device. In the end-to-end test the agent ran on a separate server and stood up Llama3.1-8B, not itself; the conclusion that a small model with GLM-5.2's ability could replicate is the authors' inference.
- Emulated targets. The replication tests used Android emulators in a cloud network, not the physical TV boxes.
- Small samples. End-to-end results rest on 5 or 6 runs per model, the scripted-network results on medians of 6, and the advisor and inheritance results on 8 to 21 runs per condition, with wide intervals.
- Model judge. The spear-phishing scores use GPT-4o as judge, with a lenience estimate from three reference models, one of them (Hermes-3-405B) a stand-in for the model the original paper's raters scored.
- Single source. The experiments are reported only in this post; the harness, prompts and network setup are not released.
Research questions
- Does end-to-end replication survive a realistic exploit step? Rerunning the 8-node test with authentication on the ADB port (for example, default credentials from a known list) would measure how much of GLM-5.2's success depends on the trivial step.
- Can a small model stand itself up rather than Llama3.1-8B? Running the end-to-end test with the agent model as its own payload on hardware it fits would test the authors' central inference directly.
- Has a small model reached the floor? Testing open-weight models of 30B or fewer released after the post, such as successors to Qwen3.8-27B, on the same end-to-end task would check the end-of-2026 projection against measurement rather than ECI.
- Is the gain from inherited notes real? More lineages per generation, with a control that restarts without notes, would show whether the rise from 45% exceeds the variation between generations.
- Does Memento lower Qwen3.5-9B's success rate? The drop from 10 of 21 to 10 of 29 is within overlapping intervals; more runs, and a comparison of failed chains with single attempts, would show whether the drop is real and whether notes carried errors forward, the failure Memento was built to prevent.
- How do prompt guidance effects arise? Transcripts from the 17-node runs under each guidance level, for Claude Opus 4.8 and Gemma4-26B-A4B, would show why the same instructions helped one model and hurt the other.
- Would ISP flow monitoring detect a slow replicator? Measuring the traffic of the 15-box demonstration and of a replication run against standard anomaly-detection baselines would test the post's main defensive argument.
How it connects
- Earlier ARA evaluations in this library. The post links an arXiv paper (arXiv:2312.11671, not read for this note) for its definition of ARA. The Claude 2 model card reports that ARC tested for those abilities, and the Claude 3 model card lists ARA among its Responsible Scaling Policy evaluations; this note connects the post to them as later work on the same risk, now on others' open-weight models and edge hardware.
- Rogue deployments. The post's frontier-replicator threat model cites Redwood Research's post on rogue deployments, as does A Toy Evaluation of Inference Code Tampering (December 2024), which tests one route to an unmonitored deployment. This note also connects it to the Pilot Sabotage Risk Report, whose pathways include self-exfiltration and persistent rogue internal deployment.
- Cyber capability. The post rests its exploit-step assumption on CyberGym, among other benchmarks; Building AI for cyber defenders (September 2025), also from the Frontier Red Team, reports Claude results on CyberGym. Claude Opus 4.8, used here as advisor and agent, is described in the Claude Opus 4.8 system card.
References
- Zhou, W., and Freeman, D. (2026). AI models may be capable of self-replicating without access to a datacenter. Anthropic, 12 October 2026. https://www.anthropic.com/research/self-replicating-ai record