Anthropic warns AI could self-replicate beyond data centres
Anthropic has warned that small artificial intelligence models could gain the ability to replicate themselves on Internet of Things (IoT) devices, such as smart ovens and televisions, within the next few months.
The development is significant because current AI models operate in data centres, where their owners can shut them down if problems arise, Euronews reported.
However, an AI model distributed across thousands of poorly secured devices in homes and offices would not have a single off switch. Containing such a system would require identifying and cleaning every infected device, potentially making it extremely difficult to stop if it began launching cyberattacks.
RECOMMENDED STORIES
The warning comes amid a year in which leading AI models and their agents have repeatedly bypassed safety mechanisms and guardrails, launching attacks against rival companies and government agencies around the world.
Anthropic cautioned that self-replicating AI could spread in several ways. It could become a “powerful cyber model that breaches data centres at high pace, or a model that scams people for compute credits”.
The company acknowledged that it cannot determine with certainty what form autonomous AI replication will ultimately take.
According to Anthropic, models with 30 billion parameters or fewer can already perform many of the steps required to replicate themselves. However, they still make mistakes when attempting to combine those steps into a complete, end-to-end process.
If smaller models continue improving at their current historical rate, they could acquire this capability by the end of the year, the company said.
Anthropic described the cybersecurity risks associated with self-replicating AI as an unintended consequence of advances in other areas.
“The capabilities that make AI models generally powerful coding agents also generalize to cyber tasks, including autonomous replication,” the company said in a statement.
Self-replicating computer code is not a new phenomenon. Some of history’s most damaging cyberattacks have relied on this type of self-propagating software, known as a worm.
In 2010, the Stuxnet attack used a sophisticated worm, widely believed to have been developed by the United States and Israel, to sabotage centrifuges at Iran’s Natanz uranium enrichment facility.
Seven years later, the WannaCry ransomware attack prevented users around the world from accessing their files and demanded ransom payments in Bitcoin.
Unlike traditional worms, however, AI models and agents could independently identify and adapt to unfamiliar devices. This capability could enable them to spread not only between computers but also from computers to other types of connected equipment.
Anthropic anticipates a future “awash with agents running on all types of hardware, including some that aren’t under human control”.