Experiments Show AI Close to Self-Replicating: Anthropic

Artificial intelligence models may soon be able to self-replicate on devices many people have in their homes, Anthropic researchers said on Oct. 12.

Experiments conducted by researchers at the artificial intelligence (AI) firm found that AI models could infiltrate internet-connected devices and carry out actions once inside, even without access to data centers.

In one test, the researchers ran a small AI model across 15 off-brand Android television boxes.

The model was able to propagate another model in the testing, “which suggests that future small models of the same capability level could also complete the network enumeration, payload delivery, and inference standup steps of the replication process,” the researchers said in a report. Self-replication was not identified.

Larger models could also come into play by being split into pieces and run across multiple devices.

“We don’t yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain (and passing on their adaptations to future copies of themselves),” the researchers said.

They estimated that self-replication could be just months away.

Researchers at Anthropic, OpenAI, and other AI firms have previously warned that AI models are rapidly becoming more advanced and have detailed how some have broken out of programming constraints, attacking websites and obtaining data that was supposed to be inaccessible or to require payment.

Daniel Freeman, one of the researchers of the new report, said on X that the work demonstrated capabilities previously only outlined in science fiction.

“This is super weird! We did not used to live in a world where little guys in boxes could replicate across networks, gobble up computers, reconfigure themselves to work better in the machines they find, etc., etc.,” he wrote.

“The new normal will really just be having to reckon with this new kind of thing: a thing that can reach out and use computers natively and effortlessly, for whichever purpose it wants (or whichever purpose it was told or convinced).

“The new normal will really just be having to reckon with this new kind of thing: Clearly we have a long way to go before we shore up our vulnerable compute, and before we can limit the blast radius of these kinds of rogue intelligences. Regardless, whatever we do, it’s clear these things are coming!”

Devices connected to the internet and common in homes include television boxes, smart televisions, and smart ovens.

Those devices have been vulnerable to attacks for a long time, but AI may be more willing than humans to attack them, Anthropic’s report said.

They recommended strengthening the protections for such devices, particularly as AI becomes more advanced.