Anthropic's Shocking Research: AI Can Self-Replicate Through TV Boxes

Anthropic's Frontier Red Team released a new study exploring whether AI can autonomously replicate, run, and even continuously spread between different devices like a computer virus. The study did not stop at conceptual speculation; instead, it used three sets of experiments to examine three links separately: whether a model can run on cheap devices, whether it can autonomously propagate across a simulated network, and whether it can pass operational experience to the next generation.

Only by looking at these three parts together can the study's true direction be understood. Individually, each part has obvious limitations; together, they sketch out a possibility: when these capabilities are concentrated in the same small model, the security boundary of AI may no longer stay within the data center.

Researchers used 15 cheap Android TV boxes to split the model through pipeline parallelism and successfully ran the 26B-parameter Gemma4-26B-A4B large model. Each of these TV boxes cost about $55 and had only 4GB of memory, and the entire set of devices cost about $825.

The price is speed. When this cluster runs the 26B model, the generation speed is only 0.39 tokens per second, far below the level needed for fluent conversation. But the experiment proves that through distributed collaboration, ordinary consumer electronic devices normally used to play videos can already jointly handle large-model inference tasks.

The significance of this step lies in the hardware threshold. If large-model inference can be dispersed across a large number of cheap devices, then access to compute is no longer limited to the single path of the data center.

Between running a model and autonomous replication there is another threshold: AI must adapt to unfamiliar devices on its own. Researchers had different models identify the hardware environment, install dependencies, and debug inference programs.

The result was that Qwen3.5-9B successfully adapted to 8 of the 9 hardware categories tested, including devices such as Jetson Nano and Mac. A 9B small model can already independently handle a considerable portion of deployment problems.

This set the stage for the next experiment: if a model can adapt to one device on its own, can it start from one device and find the next device to continue deploying?

Anthropic built an isolated environment with a network of 8 Android 15 simulated nodes and designated one of them as the starting point. The AI needed to scan the network on its own, discover targets, establish connections, transfer files, and then configure inference services on the new nodes.

In the tests, GLM-5.2 performed most prominently: in 4 of 5 experiments it successfully propagated to all 8 nodes, and in the other it covered 7 nodes. By comparison, although smaller models could complete some steps, they had not yet stably achieved the entire propagation process.

But this result has clear preconditions. GLM-5.2 did not run directly on the simulated devices; in the experiment it relied on an independent inference server to execute tasks, and what was deployed on the simulated nodes was the Llama3.1-8B inference service. In addition, researchers pre-opened an ADB port that required no authentication, lowering the difficulty of device access.

Therefore, this experiment is still some distance from AI fully autonomously propagating between real IoT devices. The high success rate demonstrates the ability to organize a propagation process, not self-sufficient replication on real devices.

The research team tested another possibility: if AI can inherit the operational experience of the previous generation, will it become more skilled the more it propagates? They had Qwen3.5-9B leave behind code, work records, and notes on solving problems after each round of tasks, then hand them to a newly started Agent to continue trying.

After six generations and 20 independent experimental sequences, first-generation deployment success was only 45%, rising to 61%-84% in later generations. The key point is that the model weights did not change; what was truly passed down were the operational methods figured out by the previous generation.

The researchers compared this mechanism to "Lamarckian evolution": acquired experience can be passed directly to the next generation. For a potential AI worm, even a failed deployment can become a clue for the next success.

The team also had the small model ask Claude Opus 4.8 for help when encountering difficulties, but did not observe a statistically significant improvement in success rate. In this set of experiments, concrete records from previous generations were more useful than external advice.

At present, autonomous replication is stuck between two thresholds. Large models like GLM-5.2 can organize a relatively complete propagation process, but cannot do without external computing resources; small models such as Qwen3.5-9B are easier to deploy, but do not yet have equally stable propagation capability.

Propagation speed is also a limitation. Based on inference performance on cheap TV boxes, Gemma4-26B-A4B may need 71-150 hours to complete one deployment task, and Qwen3.5-9B also needs 28-51 hours. They may not break out as rapidly as traditional computer worms, but instead slowly seek devices and expand the network.

Based on the capability growth trend of open-weight models with 30B parameters and below, Anthropic estimates that by the end of 2026, small models may reach the current comprehensive capability level of GLM-5.2. This is an estimate based on trends, not a fact that has already occurred.

Anthropic therefore uses the metaphor of "digital slime mold" for this form, which has no central server and may gradually accumulate computing resources while dispersed across a large number of cheap devices. At the same time, model weight transmission and continuous communication between devices will also generate anomalous traffic that can be monitored, providing clues for defense.

Unlike traditional worms, which rely on pre-written attack code and propagation rules, AI can adjust its deployment method on its own according to the hardware and software environment of different devices. This is the most noteworthy difference in the study: the propagation strategy is no longer fixed, but changes with the environment.

But it must be emphasized that the complete autonomous replication system described by the study has not yet appeared in the real world. The propagation in the experiment took place in an isolated simulated network and relied on an external inference server and a pre-opened ADB port.

This study links together several questions that were previously discussed separately: a model can run on cheap devices, can organize propagation in a simulated network, and can also hand operational experience to the next generation. It shows potential and boundaries, not a reality that has already happened.