Anthropic sounds the alarm over “self-replicating agentic worms”

On 12 October 2026, two researchers from Anthropic, William Zhou and Daniel Freeman, published a study entitled ‘AI models may be capable of replicating themselves without the need for a data centre’. Their theory is that computer worms – malware capable of spreading autonomously across networks without the need for further human intervention – powered by artificial intelligence, could demonstrate an unprecedented capacity for adaptation, posing a massive threat to global digital security. This is particularly concerning because, as the two researchers note, these agent-based worms would not require significant computational power – and would therefore not necessarily need to rely on data centre GPUs – to continue their spread and destructive work; instead, they could exploit the small Internet of Things (IoT) devices that we often use ourselves in our homes as channels of propagation. Webcams, routers, TVs, even certain fridges or vacuum cleaners.

This would result in a cyber butterfly effect, whereby a webcam in Paris could infect a nuclear power station in Illinois, or a domestic appliance in Berlin could attack a data centre campus in Shanghai. The alarm is real and serious, but as is often the case with Anthropic, the tone adopted and the nuances with which the analysis is coloured indulge in a certain degree of doomsday thinking that ultimately obscures a clear distinction between what should be taken seriously and what appears to be mere storytelling. For example, the assertion ‘we are just a few months away from’ should be taken with a pinch of salt, because there is no reliable predictive model that can truly quantify the timeframe. Even the ‘anthropomorphising’ conclusion, labelled ‘signs of life’ – as if this radical adaptation were to give rise to the development of a will attributable to the worm agents – lacks any objective basis. Let us, however, stick to reality, and reality is disturbing enough without the need for funereal trappings straight out of a medieval apocalypse. The researchers at Anthropic were not the only ones – nor, in fact, the first – to reach such conclusions, and this gets to the heart of the matter: in June 2026, the study ‘AI Agents Enable Adaptive Computer Worms’, by Jonas Guan and others, had already identified extraordinary capabilities for propagation and functional adaptation in experimental environments. This study also highlighted an even more disturbing finding: the worms utilised compromised computational resources, thereby drastically reducing the cost of the attacks. Put another way: attackers can easily bear infinitely lower financial burdens. It is well known that self-replicating worms have existed for years; the most striking example in recent history is undoubtedly WannaCry, dating back to 2017, but we are now witnessing a quantum leap in both scale and sophistication. This is because worms governed by artificial intelligence, unlike ‘classic’ worms that are standardised around a set of commands provided by the hacker, can adapt their behaviour to evade security systems, adopt diverse strategies, and analyse and identify further vulnerabilities.

Attacks on small, domestic IoT systems facilitate the spread of malware, relying on what appear to be insignificant levels of computing power. This is a significant issue, as it demands a shift in perspective regarding digital security: no longer focusing solely on critical infrastructure, but looking down to the micro-physical level, in the knowledge that the attack surface is now everywhere, even inside a fridge. At present, the most plausible risk scenario to envisage is that of widespread infection of IoT-based domestic environments, which could be exploited to compromise businesses or institutional systems. It is more difficult to envisage the spread to critical infrastructure, even though the danger posed by agent-based worms lies in the fact that they have no epicentre, no switch to turn them off – neither human nor synthetic – and therefore continue to propagate, in a labyrinthine manner, seeking escape routes and access to targets identified as and when they arise. Naturally, they would encounter a series of defensive factors—equally agent-based and of a technical nature—such as power constraints, protected environments and memory exhaustion. The real crisis—if one were to indulge in catastrophic scenarios—would lie in the possibility that these worms might not only propagate autonomously but also begin to modify their own code. In that case, they would risk adapting to the defences, thereby overcoming them more easily.