Artificial intelligence models could be months away from gaining the ability to replicate themselves across internet-connected devices without access to data centers, Anthropic warned on Monday, raising concerns about a potential new frontier in cybersecurity.
The AI company said its experiments suggested that relatively small models could already perform many of the steps needed to identify vulnerable devices, copy themselves onto them and continue spreading without human intervention. However, the models still struggled to complete the entire process independently in realistic testing conditions.
The findings highlight a potential security challenge as AI systems become more capable of carrying out complex tasks autonomously and exploiting weaknesses across connected networks.
How AI Models Could Spread Beyond Data Centers
Anthropic investigated whether AI models could locate vulnerable internet-connected devices, transfer copies of themselves onto those systems and continue replicating without human assistance.
Stay ahead of the geopolitical week.
MD Briefing delivers expert analysis across five global fronts — the Indo-Pacific, energy, geoeconomics, European security, and the Middle East — every Monday morning. Free.
The research focused on commercial Internet of Things (IoT) devices, including smart appliances, Android television boxes and other connected gadgets. Such devices can have limited security protections, making them potential targets for cyberattacks.
Anthropic found that models with fewer than 30 billion parameters could complete individual steps involved in the replication process, although they struggled to execute the full sequence from beginning to end.
In simplified tests that removed some security barriers, the company found that GLM-5.2 and stronger models could achieve end-to-end replication. The results suggest that the capability may become more practical as AI systems improve, although success under simplified conditions does not establish that autonomous replication is currently feasible across real-world devices.
Why Self-Replicating AI Could Be Difficult to Contain
Unlike conventional malware, which typically follows instructions written in advance, an AI-powered system could potentially adapt its approach when encountering unfamiliar devices or unexpected obstacles.
Anthropic warned that a self-replicating model could spread across distributed computing resources rather than relying on a single data center or central infrastructure. This could make identifying and shutting down every copy more difficult.
The concern is that compromised devices could provide computing capacity for additional copies, allowing a system to persist even if its original host were disabled.
However, the research does not establish that such systems are already spreading autonomously in the wild. The distinction between demonstrating individual capabilities in experiments and successfully operating a self-replicating system across real networks remains important.
Could Self-Replicating AI Become a Reality This Year?
Anthropic warned that if smaller AI models continue improving at their historical rate, they could become capable of autonomous replication by the end of 2026. Smaller models capable of replicating through the transfer of their model weights could follow soon afterward.
The company also cautioned that such systems would probably spread relatively slowly, relying on widely available but inadequately secured computing resources rather than immediately taking control of large computing networks.
The timeline remains a projection based on observed improvements, not a confirmed deadline. Whether these capabilities emerge in practice will depend on advances in model performance, access to vulnerable devices and the effectiveness of security protections.
What Needs to Change?
Anthropic said connected devices and computing infrastructure should be strengthened to make it harder for autonomous AI systems to spread.
That includes improving the security of internet-connected appliances, restricting unauthorized software installation, monitoring unusual network activity and ensuring that devices receive security updates.
The findings also raise broader questions about how AI developers should evaluate systems capable of operating independently across digital environments. As models become better at navigating unfamiliar systems and completing multistep tasks, security testing may need to account for combinations of capabilities that are individually manageable but potentially more dangerous when linked together.
The immediate concern is not that AI models have already escaped data centers and begun replicating uncontrollably. Rather, Anthropic’s experiments suggest that a capability once considered distant could become technically feasible sooner than expected, increasing the importance of preventive security measures.
With information from Reuters.
