441e85161ClaimAnthropic warns small AI models may be months from self-replicating on smart TVs and other IoT devicesAgreeDisagree
441e85161ClaimThe first AI disaster will be a self-replicating AI worm, not a rogue superintelligenceAgreeDisagree
441e85161ClaimSmall open-weight AI models should be restricted because they could self-replicateAgreeDisagree
441e85161ClaimAnthropic estimated in October 2026 that about 2 million infected smart TVs could run roughly 133,000 self-replicating AI model copies producing about 80,000 tokens per secondAgreeDisagree
441e85161ClaimUnlike traditional botnets such as the Kimwolf IoT botnet, a self-replicating AI botnet would not be tied to one hardware or software family, according to AnthropicAgreeDisagree
441e85161ClaimIn Anthropic's October 2026 simplified self-replication tests, Z.ai's GLM-5.2 was the only open-weight AI model to replicate end to end, spreading to all 8 nodes of a test network in 4 of 5 trialsAgreeDisagree
441e85161ClaimAnthropic called its October 2026 AI self-replication results an upper bound because its test devices had an unauthenticated Android Debug Bridge port as the entry pointAgreeDisagree
441e85161ClaimWhen Anthropic scripted the inference setup for it, a 3-billion-parameter AI model self-replicated to about half of a 17-node test network in Anthropic's October 2026 experimentsAgreeDisagree
441e85161ClaimAnthropic warned in October 2026 that an AI model self-replicating across IoT devices would be hard to shut down because it would have no central point to switch offAgreeDisagree
441e85161ClaimIn Anthropic's October 2026 self-replication tests, open-weight AI models of 30 billion parameters or fewer completed many individual replication steps but made mistakes chaining them togetherAgreeDisagree
441e85161ClaimTo show AI models can run on cheap consumer devices, Anthropic ran Google's Gemma4-26B-A4B model across 15 off-brand Android TV boxes costing about $55 each, at 0.39 tokens per secondAgreeDisagree
441e85161ClaimAnthropic's Frontier Red Team estimated in October 2026 that open-weight AI models of 30 billion parameters or fewer could self-replicate end to end by the end of 2026 if they keep improving at their historical rateAgreeDisagree
441e85161ClaimIn Anthropic's October 2026 self-replication tests, a 9-billion-parameter open-weight AI model set up its own inference on 8 of 9 hardware types, from Nvidia Jetson Nanos to MacsAgreeDisagree
441e85161ClaimAdvice from Anthropic's Claude Opus 4.8 did not significantly improve how often Alibaba's small Qwen3.5-9B model set up its own inference in Anthropic's October 2026 self-replication testsAgreeDisagree
441e85161ClaimAnthropic's Frontier Red Team reported on October 12, 2026 that small AI models may be months away from being able to self-replicate on commercial IoT devicesAgreeDisagree
441e85161ClaimSelf-replicating AI models pose a greater cybersecurity threat than traditional malwareAgreeDisagree
441e85161ClaimA rogue AI agent swarm could seize much of the internet within 6 to 12 monthsAgreeDisagree