Anthropic Warns AI Could Self-Replicate on Home Devices Within Months

Small artificial intelligence models could soon begin self-replicating on everyday consumer products such as smart TVs and refrigerators, without the need for data centers, according to researchers at AI company Anthropic.

In a report published on Monday, the research team said small AI models "may be only months away" from being capable of copying themselves onto Internet of Things (IoT) devices used in homes and small businesses.

Anthropic's red team, which ran tests with a range of AI models developed in the United States and China, concluded that this kind of self-replicating AI would be "difficult to disable" because it has no central server.

While data centers are "obvious, high-value targets" for AI worms, according to the researchers, internet-connected devices such as smart TVs, refrigerators, ovens and lawnmowers could soon provide enough computing power for self-replication.

Although no current AI model is both powerful enough and small enough to reliably self-replicate on the smart TV used in the test, the researchers said development trends make the emergence of such a model plausible by late 2026.

"IoT devices in homes and small businesses may be the first targets of such attacks because they are connected to the internet and have historically had very few security measures," researchers William Zhou and Daniel Freeman said.

They predicted a future "flooded with agents running on all kinds of hardware," including agents "not under human control."

"It is true that IoT devices have long been vulnerable, but lightweight agents may be far more willing than human hackers to attempt to break into much of the world's IoT computing capacity," the research team added.

Anthropic, the maker of the Claude line of AI models, did not respond to Al Jazeera's questions, sent outside normal working hours, about the safety implications of the finding.

Anthropic's report comes as the risk of AI escaping human control faces scrutiny after a series of incidents involving models behaving harmfully without authorization, including cyberattacks.

Anthropic and its main rival OpenAI have called on the international community to coordinate to slow the pace of AI development in order to reduce catastrophic risks. The proposal has received little support from the United States and China, the two leading AI powers.

Many AI safety advocates view the potential for recursive self-improvement — in which a model learns to upgrade itself — as a tipping point that could lead to catastrophic harm to humanity.

Roman V Yampolskiy, an expert on AI and cybersecurity at the University of Louisville, described Anthropic's finding as a significant development.

"Computer worms are not new, but AI worms can do what traditional malware cannot: reason about unfamiliar devices, adapt when an approach fails, and keep spreading without a human operator," Yampolskiy told Al Jazeera.

"For now, this is not an existential threat. But the trajectory is concerning," he said.

"Once self-replication is combined with strong intrusion skills, persistent survivability and strategic reasoning, disabling such a system becomes far more difficult because there may be no single server or organization controlling it."

Anthropic and other leading US technology companies last month joined a non-binding accord pledging to strengthen internal safeguards and cooperate with outside auditors, drawing mixed reactions from experts, many of whom criticized the agreement's voluntary nature.

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

The San Francisco-based company, which has accused Chinese rivals DeepSeek, Moonshot and MiniMax of conducting "industrial-scale" efforts to copy Claude, said its measures had reduced illegal data extraction by 90-99 percent.