AI Tech36氪

Anthropic Shocking Study: AI Could Self-Replicate Through TV Boxes

36氪Oct 14, 20262 min read

AI Summary

Anthropic's Frontier Red Team published research examining whether AI could replicate, run, and spread across devices autonomously, like a computer virus. The team ran a 26B-parameter model across 15 low-cost Android TV boxes and found that small models can increasingly adapt to unfamiliar hardware. The experiments were conducted in isolated environments, not real-world networks.

AI Insights

Anthropic's research shows that small open-weight models are gaining the ability to adapt to unfamiliar hardware, deploy inference environments autonomously, and pass on operational experience across generations, which narrows the gap between AI capabilities and self-replicating software. The industry implication is that AI security must extend beyond data centers to cover consumer IoT devices and distributed compute, prompting stronger network monitoring, device hardening, and evaluation of autonomous replication risk as a standard part of model safety testing. Because the experiments ran in isolated environments with pre-opened, unauthenticated ports, the findings signal future risk trajectories rather than an existing threat, making proactive governance and benchmarking of replication capabilities especially timely before 2026 year-end.

Analysis written by AI from the original article. Check it against the source.

Translated text

AI may be nearing a potentially alarming capability: replicating itself. Anthropic's Frontier Red Team has published new research examining whether AI could, like a computer virus, autonomously replicate, run, and spread across devices. The researchers argue that AI could run on smart TVs, TV boxes, and smart ovens rather than relying on expensive data centers. In the experiment, the team connected 15 Android TV boxes costing about $55 each (4GB RAM, roughly $825 total) and ran the Gemma4-26B-A4B model split across them using pipeline parallelism. Generation was slow at 0.39 tokens per second, but distributed collaboration made large-model inference possible. Next, the researchers tested whether AI could identify unfamiliar hardware, install dependencies, and debug inference programs on its own. Qwen3.5-9B adapted to 8 of the 9 hardware types tested. In an isolated network of 8 Android 15 emulator nodes, GLM-5.2 spread to all eight nodes in 4 of 5 runs and to seven nodes in the remaining run. However, the experiment relied on an external inference server, the ADB ports were left open without authentication, and the nodes ran Llama3.1-8B, so it remains some distance from fully autonomous spread across real IoT devices. Another finding concerns transfer of experience across generations. When Qwen3.5-9B left code, work logs, and problem-solving notes for the next agent, the deployment success rate rose from 45% in the first generation to 61–84% across six generations and 20 independent experiment sequences. The model weights did not change; only the methods discovered by earlier generations were passed on. The researchers likened this to 'Lamarckian evolution.' Based on capability trends of open-weight models with 30B parameters or fewer, Anthropic estimates that by the end of 2026 small models could reach the overall capability level of today's GLM-5.2. Spread could be slow, so Anthropic called it a 'digital slime mold': distributed across many cheap devices without a central server, gradually accumulating compute. No fully autonomous replication system of the kind described has been found in the real world. Still, the study connects issues previously discussed separately: running on low-cost devices, spreading across simulated networks, and passing experience between generations. If these capabilities converge in a single small model, AI safety boundaries may no longer stop at the data center.

Related News

AI Tech36氪

Google Awakens to Physics: Gemini 4 Argon Hands-On Tests Pour In

Google's Gemini 4 Argon has launched and reportedly competes closely with GPT-6 Astra and Claude Opus 5.5 on benchmarks, at about one-fifth the price and several times faster. Hands-on tests showed strengths in 3D generation, physics-intuitive visualization, and scientific reasoning, though its Chinese-language narration sounded awkward.

AI Tech36氪

Proof Explosion Reshapes Mathematics: AI Pushes Terence Tao Toward 'Math 2.0'

On October 9, Fields Medalist Terence Tao gave a Caltech lecture, retitled 'Math 2.0' as AI advances rapidly. He argues mathematics has shifted from an era of scarce proofs to one of abundant proofs. He warns that producing answers without human understanding is a dangerous path.

AI Tech虎嗅

Hands-On Test of Gemini 4 Argon: Google's Stumbling Last-Ditch Effort

Google has made Gemini 4 Argon generally available after first announcing it on October 1 for a security program. Following a year of talent departures and delayed model launches, the model is widely seen as a last-ditch effort for Google. Hands-on testing suggests it will struggle to challenge Claude Opus 5.5 or GPT-6 Astra.