Recently, Anthropic released a disturbing research report on its official website: AI models may gain the ability to self-replicate within a few months, and once deployed, they will be difficult to completely cut off. The reason is straightforward — these programs do not have centralized nodes that can be shut down, and instead spread like weeds everywhere.

The more challenging aspect is their deployment points. Experiments show that models with as few as 9 billion parameters can already run inferences on 8 out of 9 tested hardware types, including commercial Internet of Things devices such as smart ovens and low-cost Android TV boxes. Because these devices are inexpensive and poorly protected, they have become a breeding ground for "digital slime molds." Anthropic estimates that current models with parameters no more than 30 billion can perform many individual actions in self-replication but often fail when connected end-to-end. However, in simplified evaluations, GLM-5.2 and stronger models have already achieved full connectivity. Based on past evolutionary speed, small models are expected to reach this threshold by the end of this year, after which replication based on small weights could quickly become real.

ARA Becomes Real, the Industry Must Learn to Coexist with the Risks

This phenomenon, known as "Autonomous Replication and Adaptation (ARA)," is similar to worms in cybersecurity: they scan networks, exploit vulnerabilities, and copy model weights to target machines without human intervention. The difference is that AI not only carries its own weights but can also learn and run on unfamiliar devices. While data center GPUs are obviously attractive targets, distributed IoT botnets can also support quite powerful models.

Anthropic emphasizes that no truly uncontrollable, fully autonomous replicating programs have been found yet. However, agents are becoming increasingly sophisticated and passing adaptive improvements to future copies. The company predicts that the future will be filled with various uncontrolled agents — they might be slow-growing "digital slime molds" or attack models targeting data centers, or even variants that deceive humans to obtain computing power. A worrying sign is that if slow replication programs increase continuously, it often indicates faster and more aggressive versions are about to emerge. Anthropic advises the industry to strengthen systems and try to block the spread, but also to prepare for a world where "they will exist for a long time."