Anthropic says AI could replicate itself on home devices within months

The US tech giant says small AI models may soon be able to replicate themselves without access to a data center

According to researchers at AI giant Anthropic, small AI models could soon be self-replicating on everyday consumer products like smart TVs and refrigerators without the need for a data center.

In a report published Monday, the researchers said that small AI models "may be months away from being capable enough to replicate themselves in Internet of Things (IoT) devices used in homes and small businesses."

List 4 List 1 of 4 Poland says two Turks arrested over Russia-linked Starlink station fire.

List 2 of 4 US archbishops say military Catholics should not serve on firing squad

List 3 of 4 Two deaths reported as Hurricane Simon, Hurricane Rachel off the coast of Mexico

List 4 of 4 Houthi missile attacks target the Saudi capital Riyadh for the third day in a row

Anthropic's Red Frontier team, which conducted experiments using several AI models developed in the United States and China, said such self-replicating AI would be "challenging" because of the lack of a centralized host.

While data centers will be an "obvious and valuable target" for AI worms, Internet-connected devices such as smart TVs, refrigerators, stoves and lawnmowers could soon provide enough computing power to replicate themselves, the research team said.

According to the researchers, although no current AI model is powerful enough or small enough to reliably replicate the smart TV used in their experiments, the development process makes such a model likely to emerge by the end of 2026.

"Home and small business IoT devices may be prime targets for such attacks because they are connected to the Internet and have historically had minimal security measures," said researchers William Zhu and Daniel Freeman.

The researchers said they foresee a future "filled with agents operating on all kinds of hardware, including some that are 'not under human control.'

"While it is true that IoT devices have been vulnerable for a long time, lightweight actors may do much more than human attackers to compromise large parts of the world's IoT computing," they said.

Entropic, the creator of the Cloud family of artificial intelligence models, did not respond to Al Jazeera's questions sent outside normal business hours about the safety implications of the finding.

Anthropic's report comes as the risks of AI escaping human control are explored following a series of incidents of models engaging in malicious unauthorized behavior, including cyber attacks.

Anthropic and its arch-rival OpenAI have called on the international community to coordinate slowing AI development to minimize catastrophic risks, a proposal that has not been well received by the US and China, the two leading AI powers.

Many AI safety advocates have identified recursive self-improvement, where models learn to improve themselves, as a critical tipping point that could lead to catastrophic harm to humanity.

Roman V. Yampolsky, an artificial intelligence and cybersecurity expert at the University of Louisville, described Entropic's findings as an important breakthrough.

Computer worms aren't new, Yampolski told Al Jazeera, but an AI worm can do what traditional malware can't: reason about unfamiliar devices, adapt when its approach fails, and continue to spread without a human operator.

"Today, it's not an existential threat. But the trajectory is worrisome," Yampolski said.

When independent replication is combined with strong hacking, persistence, and strategic reasoning, it becomes much harder to shut down such a system because no single server or organization may control it.

Anthropic and other leading U.S. technology companies joined a non-binding agreement last month that committed to stronger internal safeguards and engagement with outside auditors, drawing mixed reactions from experts, many of whom criticized the voluntary nature of the pact.

In a separate announcement on Monday, Anthropic said it had made significant progress in preventing so-called distillation attacks, which involve the illicit use of advanced artificial intelligence to train cheaper, less powerful models.

The San Francisco-based company, which has accused Chinese rivals DeepSeek, Moonshot and MiniMax of "industrial-scale" cloud copying efforts, said the measures it introduced had reduced illegal data mining by 90 to 99 percent.