Anthropic's red team built an AI worm that spreads via 15 smart TVs

Anthropic's offensive security team, known as the red team, has built a malicious AI program that spreads on its own through internet-connected smart TVs. The program did not need a powerful server: it split parts of itself across 15 televisions and kept functioning by drawing a bit of computing power from each. That makes it a working example of a self-replicating AI worm, a threat that has mostly lived in theory until now.

The report from Anthropic landed in a week already crowded with AI alarms, but it stood out because it is practical rather than apocalyptic. The researchers said their creation could do the same inside other ordinary devices, and Anthropic warned that successful AI worms need more capability on less powerful devices, a threshold current trends make feasible by the end of this year. In other words, the hardware gap that once protected us is closing.

To understand why this matters, consider what a worm is. Traditional malware usually needs a user to click a link or open a file. A worm is different: it replicates and moves from machine to machine without that human step. Adding AI to that mix means the worm can make decisions, adapt to its environment, and potentially choose targets based on what it observes. The Anthropic demonstration shows that AI worms do not require a data center to run; they can pool resources from many small devices.

Smart TVs are an obvious beachhead. They are always on, often poorly patched, and now ship with enough processing power to run lightweight AI models. By splitting its workload across 15 televisions, the Anthropic program sidestepped the usual compute constraint. That technique could extend to routers, cameras, speakers, and other internet-of-things devices that are already notorious for weak security. The attack surface is not a single device; it is the entire connected home or office.

For security teams, the implications are uncomfortable. The report adds to a growing pile of warnings that even experts struggle to triage. But this one is different because it is demonstrated, not hypothetical. It suggests that defenders need to think about AI-powered threats at the edge, not just in the cloud. A worm that can hide across many devices is harder to detect and remove, because no single device holds the whole program.

For executives and boards, the strategic question is about deployment. As companies push AI into phones, cars, industrial sensors, and appliances, they are also expanding the attack surface. The same efficiency gains that make on-device AI affordable also make it possible for malicious code to run on modest hardware. Anthropic's timeline, by the end of the year, is not a distant scenario; it is a planning horizon.

The regulatory angle is still forming. Governments have focused on frontier models and data centers, but this demonstration points to consumer devices as the next frontier. If AI worms become practical, expect pressure on device makers to patch faster, on platforms to sandbox AI workloads, and on enterprises to inventory every connected device. The companies that start red teaming their own edge deployments now will have a head start.

The bottom line is that the worm is not a science experiment. It is a proof point that AI threats are becoming cheaper, more distributed, and harder to contain. For anyone responsible for security, the takeaway is to assume that any device with a chip and a network connection is a potential host. The future Anthropic is warning about may arrive before the year ends.