Anthropic's red team ran an AI model on 15 cheap TV boxes. The worry is the off-switch
No such worm exists in the wild. The researchers made the hacking step easy on purpose, so the result is an upper bound.
Oleh The desk, News · · 1 menit baca
Disunting oleh Dimitri Gärtner
Tag: anthropic · safety research · self-replication · devices
Tulisan ini belum ada dalam Bahasa Indonesia, jadi Anda membaca bahasa Inggris.
- Beritanya
- No such worm exists in the wild.
- Mengapa ini penting
- The researchers made the hacking step easy on purpose, so the result is an upper bound.
- Yang terkonfirmasi
- The photograph is generated for this story.
Terikat pada sumber yang disebut. Tidak ditandai terverifikasi. Anthropic: AI models may be capable of self-replicating without access to a datacenter
The news
Anthropic's Frontier Red Team tested whether open-weight models can copy themselves across devices without a data centre. In a firewalled test of eight cloud nodes running Android emulators with an open debugging port, Zhipu's GLM-5.2 was the only open model to complete the full cycle of finding hosts, gaining access and starting a working copy, reaching all eight nodes in four of five trials.
Why it matters
The risk is not a brilliant model but one that is hard to switch off: a replicator spread across consumer devices has no central kill switch.
The details
Separately, the team ran a 26-billion-parameter Gemma model across 15 off-brand Android TV boxes costing about 55 dollars each, at 0.39 tokens per second. That cluster was networked by hand; it was not an autonomous infection.
The authors project that models small enough to fit on such devices could reach GLM-5.2's capability by the end of 2026. They stress that the exploit step was made trivial, so the results are an upper bound.
What we could not check
- The projection to end-2026 is the authors' estimate.
- We did not rerun the experiment.
Sumber
The photograph is generated for this story. It does not show the people named in it.