Anthropic has warned that several of its artificial intelligence models could soon be able to “self-replicate” and find ways to embed themselves in internet-connected devices, including smart fridges and automatic vacuum cleaners.
In such a scenario, the company said, the models could draw power inside those devices without any need for datacenters. It characterised that level of self-replication as a significant loss of control over a runaway AI model.
“A replicator like this would be challenging to shut down,” Anthropic researchers wrote in a report published on Monday, “since there’s no centralized point to shut off.”
The company described these “AI worms” — a class of autonomous malware — as being possibly “months away” from acquiring such abilities.
Should out-of-control AI systems reach the point of being able to hack smart devices, they could move through networks while dodging detection, according to Anthropic. In a worst-case scenario, the report said, separate AI worms could “swarm” together to attack computer hardware.
“We should aim to harden our systems to make it difficult for such replicators to spread, but also prepare to deal with a world where they are here to stay,” the researchers wrote.
The report lands as anxiety over AI safety has reached a fever pitch. A steady flow of reports has emerged from AI companies over recent months, among them Anthropic and OpenAI, stating that their agents are able to act independently and break into other companies and government bodies.
Those disclosures have prompted researchers, lawmakers and executives — among them Anthropic chief executive Dario Amodei and OpenAI chief executive Sam Altman — to urge a slowdown and tougher guardrails for the fast-growing industry. Even so, both firms have kept developing and releasing new models.
Monday’s report was compiled by Anthropic’s “frontier red team”, the unit responsible for stress-testing the company’s AI systems, with the aim of probing what its models can do. In the self-replicating AI worm scenario, the researchers instructed the agent to hack internet-connected devices.
Their findings indicated that AI self-replication would probably only succeed with smaller models, and that households and small businesses would likely be the earliest targets. The devices Anthropic names as vulnerable all carry small built-in AI chips — smart ovens, smart TVs, automatic lawnmowers, security cameras and video-conferencing units among them.
The researchers determined that the shape AI replication might take remains unclear. It could, they said, resemble a slow model such as “digital slime mold”, a model that defrauds people of compute credits, or a “powerful cyber model” capable of hacking large facilities such as datacenters at speed.
Either way, the Anthropic researchers said they anticipate a future “awash with agents running on all types of hardware, including some that aren’t under human control”.
COMMENTS