--- title: Anthropic’s red team ran AI on 15 cheap TV boxes. Then... description: Anthropic's Frontier Red Team built a 15-node cluster of cheap Android TV boxes for ~$825 and ran AI inference on it to study whether small open-weight models... canonical: https://daily.dev/posts/anthropic-s-red-team-ran-ai-on-15-cheap-tv-boxes-then-the-math-got-unsettling--jqlu7v8f6 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Anthropic’s red team ran AI on 15 cheap TV boxes. Then the math got unsettling. | daily.dev og:description: Anthropic's Frontier Red Team built a 15-node cluster of cheap Android TV boxes for ~$825 and ran AI inference on it to study whether small open-weight models... og:url: https://daily.dev/posts/anthropic-s-red-team-ran-ai-on-15-cheap-tv-boxes-then-the-math-got-unsettling--jqlu7v8f6 og:image: https://api.daily.dev/og/posts/jQlu7v8F6.png og:image:alt: Anthropic’s red team ran AI on 15 cheap TV boxes. Then the math got unsettling. og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Anthropic’s red team ran AI on 15 cheap TV boxes. Then the math got unsettling. **[The New Stack](https://daily.dev/sources/newstack)** · 6 min read · 0 upvotes · 0 comments ## Summary Anthropic's Frontier Red Team built a 15-node cluster of cheap Android TV boxes for ~$825 and ran AI inference on it to study whether small open-weight models could self-replicate across IoT hardware. Researchers William Zhou and Daniel Freeman identified four replication steps — finding a target, gaining access, running inference, and moving on — and found that GLM-5.2 successfully reached all eight nodes in a firewalled test environment in four of five trials. A key finding is 'Lamarckian inheritance': agents that read notes from prior attempts improved success rates from 45% to 61–84% across generations. The main barrier is that capable models are currently too large for cheap devices, but the Epoch Capabilities Index trend line suggests sub-30B models could reach GLM-5.2's capability level by end of 2026. Using the Kimwolf botnet's 2M-device footprint as a reference, the researchers calculated that such a network could yield ~133,000 model copies generating 80,000 tokens/second — equivalent to ~40 H100 GPUs — at zero hardware cost to the attacker. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What did Anthropic's red team experiment show about running AI models on cheap Android TV boxes? For roughly $825, Anthropic's Frontier Red Team built a 15-node cluster of off-brand Android TV boxes that runs an AI model at 0.39 tokens per second. The experiment showed that 4GB TV boxes running 4-bit weights can host a 9B model across four or more devices and a 26B model across ten or more. A single inference setup on this hardware class could take 28 to 150 hours depending on the model. _Engineers tracking AI infrastructure risks find threat research like this alongside the latest model and tooling news on daily.dev._ ### What is Lamarckian inheritance in the context of AI agents and self-replication? In Anthropic's self-replication research, Lamarckian inheritance means one agent passes notes about what it learned — such as how to handle installation failures on unfamiliar hardware — to the next agent without any retraining of the underlying model. Running Qwen3.5-9B through 20 lineages of six generations each, success rates climbed from 45% in the first generation to between 61% and 84% in later ones. _Developers building persistent agent systems can track how memory and inheritance patterns evolve across the AI agent ecosystem on daily.dev._ ### How capable would a botnet of cheap Android TV boxes running AI be, according to Anthropic's calculations? Using the Kimwolf botnet's 2 million infected Android TV boxes as a reference, dividing those devices into clusters of 15 running Gemma4-26B-A4B would yield roughly 133,000 model copies generating a combined 80,000 tokens per second. That is equivalent to no more than about 40 Nvidia H100 GPUs — modest by datacenter standards, but the hardware and operating costs would belong to the compromised device owners. _Security engineers assessing AI-driven botnet risks can follow this research thread and related threat disclosures on daily.dev._ ## Similar posts on daily.dev - [Amazon kept shutting down my tablet, so I spent $266 on four AI models to own it](https://daily.dev/posts/amazon-kept-shutting-down-my-tablet-so-i-spent-266-on-four-ai-models-to-own-it-knxdr1np3) · Lobsters · 0 upvotes · 0 comments --- Tags: [#security](https://daily.dev/tags/security), [#anthropic](https://daily.dev/tags/anthropic), [#ai-security](https://daily.dev/tags/ai-security), [#edge-computing](https://daily.dev/tags/edge-computing) [View this post on daily.dev](https://daily.dev/posts/anthropic-s-red-team-ran-ai-on-15-cheap-tv-boxes-then-the-math-got-unsettling--jqlu7v8f6) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Anthropic’s red team ran AI on 15 cheap TV boxes. Then the math got unsettling.","url":"https://daily.dev/posts/anthropic-s-red-team-ran-ai-on-15-cheap-tv-boxes-then-the-math-got-unsettling--jqlu7v8f6","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/anthropic-s-red-team-ran-ai-on-15-cheap-tv-boxes-then-the-math-got-unsettling--jqlu7v8f6"},"datePublished":"2026-10-13T16:44:36.694Z","dateModified":"2026-10-13T17:54:29.944Z","description":"Anthropic's Frontier Red Team built a 15-node cluster of cheap Android TV boxes for ~$825 and ran AI inference on it to study whether small open-weight models...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/95cc7672fa0e24cb993e2f05283e8d46?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/95cc7672fa0e24cb993e2f05283e8d46?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"The New Stack","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"The New Stack","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/newstack","url":"https://daily.dev/sources/newstack"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/anthropic-s-red-team-ran-ai-on-15-cheap-tv-boxes-then-the-math-got-unsettling--jqlu7v8f6","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"security,anthropic,ai-security,edge-computing","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The New Stack","item":"https://daily.dev/sources/newstack"},{"@type":"ListItem","position":3,"name":"Anthropic’s red team ran AI on 15 cheap TV boxes. Then the math got unsettling."}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/anthropic-s-red-team-ran-ai-on-15-cheap-tv-boxes-then-the-math-got-unsettling--jqlu7v8f6#faq","mainEntity":[{"@type":"Question","name":"What did Anthropic's red team experiment show about running AI models on cheap Android TV boxes?","acceptedAnswer":{"@type":"Answer","text":"For roughly $825, Anthropic's Frontier Red Team built a 15-node cluster of off-brand Android TV boxes that runs an AI model at 0.39 tokens per second. The experiment showed that 4GB TV boxes running 4-bit weights can host a 9B model across four or more devices and a 26B model across ten or more. A single inference setup on this hardware class could take 28 to 150 hours depending on the model. Engineers tracking AI infrastructure risks find threat research like this alongside the latest model and tooling news on daily.dev."}},{"@type":"Question","name":"What is Lamarckian inheritance in the context of AI agents and self-replication?","acceptedAnswer":{"@type":"Answer","text":"In Anthropic's self-replication research, Lamarckian inheritance means one agent passes notes about what it learned — such as how to handle installation failures on unfamiliar hardware — to the next agent without any retraining of the underlying model. Running Qwen3.5-9B through 20 lineages of six generations each, success rates climbed from 45% in the first generation to between 61% and 84% in later ones. Developers building persistent agent systems can track how memory and inheritance patterns evolve across the AI agent ecosystem on daily.dev."}},{"@type":"Question","name":"How capable would a botnet of cheap Android TV boxes running AI be, according to Anthropic's calculations?","acceptedAnswer":{"@type":"Answer","text":"Using the Kimwolf botnet's 2 million infected Android TV boxes as a reference, dividing those devices into clusters of 15 running Gemma4-26B-A4B would yield roughly 133,000 model copies generating a combined 80,000 tokens per second. That is equivalent to no more than about 40 Nvidia H100 GPUs — modest by datacenter standards, but the hardware and operating costs would belong to the compromised device owners. Security engineers assessing AI-driven botnet risks can follow this research thread and related threat disclosures on daily.dev."}}]} ```