AI could self-replicate on home devices within ‘months’, Anthropic says

US tech giant says small AI models may soon be capable of self-replication without access to a data centre.

People visit an Anthropic booth at Moscone Center during the Dreamforce 2026 technology summit in San Francisco, California, US, on September 17, 2026 [Carlos Barria/Reuters]

Small artificial intelligence models could soon begin self-replicating on everyday consumer products such as smart TVs and fridges, without needing a data centre, according to researchers at AI giant Anthropic.

In a report released on Monday, the researchers said small AI models “may be months away from being capable enough” to copy themselves on internet of things (IoT) devices used in homes and small businesses.

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

list 2 of 4 US archbishop says Catholics in army should not participate in firing squad

list 3 of 4 Two deaths reported as Hurricane Simon, storm Rachel batter Mexico’s coast

list 4 of 4 Houthi missile attack targets Saudi capital Riyadh for third day in a row

Anthropic’s frontier red team, which ran experiments using multiple US- and Chinese-developed AI models, said such self-replicating AI would be “challenging to shut down” due to the lack of a centralised host.

While data centres would be “a clear, high-value target” for AI worms, internet-connected devices like smart TVs, fridges, ovens and lawnmowers could soon also provide sufficient computing power for self-replication, the research team said.

Although no current AI model proved to be both powerful and small enough to self-replicate reliably on a smart TV used in testing, the trend in development makes the emergence of such a model likely by the end of 2026, the researchers said.

“Home and small business IoT devices might be the first target for such attacks because they’re connected to the internet and, historically, have had minimal security measures in place,” researchers William Zhou and Daniel Freeman said.

The researchers said they anticipated the future would be “awash with agents running on all types of hardware”, including some that “aren’t under human control”.

“While it’s true that IoT devices have been vulnerable for a long time, lightweight agents may be much more willing than human attackers to put in the effort to compromise large fractions of the world’s IoT compute,” they said.

Anthropic, the creator of the Claude family of AI models, did not respond to Al Jazeera’s questions, sent outside regular business hours, about the finding’s safety implications.

Anthropic’s report comes as the risks of AI escaping human control are under scrutiny following a slew of incidents of models engaging in unsanctioned malign behaviour, including cyberattacks.

Anthropic and its chief rival OpenAI have called on the international community to coordinate a slowdown in AI development to minimise catastrophic risks, a suggestion that has had little buy-in from the United States 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 for humanity.

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

“Computer worms are not new, but an AI worm could do something traditional malware cannot: reason about unfamiliar devices, adapt when its approach fails, and continue spreading without a human operator,” Yampolskiy told Al Jazeera.

“Today, this is not an existential threat. But the trajectory is concerning,” Yampolskiy said.

“Once autonomous replication is combined with strong hacking, persistence, and strategic reasoning, shutting such a system down becomes much harder because there may be no single server or organisation controlling it.”

Anthropic and other leading US tech firms last month joined a non-binding accord pledging stronger internal safeguards and engagement with external auditors, drawing a mixed response from experts, many of whom criticised the voluntary nature of the pact.

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

The San Francisco-based firm, which has accused Chinese competitors DeepSeek, Moonshot, and MiniMax of “industrial-scale” efforts to copy Claude, said the measures it introduced had reduced the illicit extraction of data by 90-99 percent.

美国科技巨头表示，小型人工智能模型可能很快就能在无需访问数据中心的情况下进行自我复制。

2026 年 9 月 17 日，在美国加利福尼亚州旧金山举行的 Dreamforce 2026 科技峰会期间，人们参观了 Moscone 中心的 Anthropic 展位。[Carlos Barria/路透社]

据人工智能巨头 Anthropic 的研究人员称，小型人工智能模型可能很快就能在智能电视和冰箱等日常消费品上进行自我复制，而无需数据中心。

研究人员在周一发布的一份报告中表示，小型人工智能模型“可能还需要几个月的时间才能具备足够的能力”在家庭和小企业使用的物联网 (IoT) 设备上进行自我复制。

波兰称两名土耳其人因涉嫌与俄罗斯有关的星链基站火灾被捕（共四条记录，此为第一条）：波兰方面表示，两名土耳其人因涉嫌参与与俄罗斯有关的星链基站火灾而被捕。

美国大主教称，军队中的天主教徒不应参与枪决（共四条记录，此为第二条）。

飓风西蒙和风暴瑞秋袭击墨西哥海岸，造成两人死亡（共4条记录中的第3条）

胡塞武装连续第三天向沙特首都利雅得发动导弹袭击，这是其中的第四起。

人类学研究所的先锋红队使用多个美国和中国开发的 AI 模型进行了实验，他们表示，由于缺乏集中式主机，这种自我复制的 AI 将“难以关闭”。

研究团队表示，虽然数据中心将是人工智能蠕虫“明确的、高价值的目标”，但像智能电视、冰箱、烤箱和割草机这样的联网设备也可能很快提供足够的计算能力来进行自我复制。

研究人员表示，虽然目前还没有任何人工智能模型能够同时兼具强大的功能和足够小的体积，从而在测试中使用的智能电视上可靠地自我复制，但发展趋势表明，到 2026 年底，这种模型很可能会出现。

研究人员威廉·周和丹尼尔·弗里曼表示：“家庭和小型企业的物联网设备可能是此类攻击的首要目标，因为它们连接到互联网，而且从历史上看，它们的安全措施非常有限。”

研究人员表示，他们预计未来将“充斥着运行在各种硬件上的智能体”，其中一些智能体“不受人类控制”。

他们表示：“虽然物联网设备长期以来确实存在漏洞，但轻量级代理可能比人类攻击者更愿意投入精力来破坏全球大部分物联网计算。”

Claude 系列人工智能模型的创建者 Anthropic 公司没有回应半岛电视台在正常工作时间之外发出的关于该发现对安全影响的问题。

人类学研究所的这份报告发布之际，人工智能摆脱人类控制的风险正受到密切关注，此前发生了一系列模型参与未经授权的恶意行为（包括网络攻击）的事件。

Anthropic 及其主要竞争对手 OpenAI 呼吁国际社会协调放缓人工智能的发展，以最大限度地减少灾难性风险。然而，这一建议并未得到美国和中国这两个领先的人工智能强国的认同。

许多人工智能安全倡导者认为，递归式自我改进（即模型学习自我改进）是一个关键的转折点，可能会给人类带来灾难性的伤害。

路易斯维尔大学的人工智能和网络安全专家 Roman V Yampolskiy 将 Anthropic 的研究成果描述为一项重大进展。

“计算机蠕虫并不新鲜，但人工智能蠕虫可以做到传统恶意软件做不到的事情：对不熟悉的设备进行推理，在方法失败时进行调整，并在没有人类操作员的情况下继续传播，”扬波尔斯基告诉半岛电视台。

“目前，这还不是生存威胁。但其发展轨迹令人担忧，”扬波尔斯基说。

“一旦自主复制与强大的黑客技术、持久性和战略推理相结合，关闭这样的系统就变得更加困难，因为可能没有单个服务器或组织控制它。”

上个月，Anthropic 和其他美国领先的科技公司加入了一项不具约束力的协议，承诺加强内部保障措施并与外部审计师合作，这引起了专家们的不同反应，许多专家批评该协议的自愿性质。

Anthropic 在周一发布的另一份声明中表示，该公司在防止所谓的“蒸馏攻击”方面取得了实质性进展。所谓“蒸馏攻击”是指未经授权使用尖端人工智能来训练更便宜、能力更弱的模型。

这家总部位于旧金山的公司指责中国竞争对手深潜科技、登月科技和MiniMax“大规模”抄袭Claude，并表示其采取的措施已将非法数据提取减少了90%至99%。