Anthropic, the US artificial intelligence company, has published a report from its offensive research team describing a demonstrable, non‑apocalyptic threat: a self‑replicating malicious program that autonomously spread across internet‑connected televisions. The team, known internally as a red team, said the prototype was able to divide its components across multiple smart TVs and continue to operate by drawing a small amount of compute from each device.
The company framed the finding as a practical security concern rather than a sensational prediction. While many warnings about AI focus on hypothetical scenarios or human‑like behavior, Anthropic’s report centers on a technical capability that already exists in a rudimentary form and could become more dangerous as models and devices evolve.
Anthropic highlighted a key technical barrier for such ‘AI worms’: running advanced machine‑learning models typically requires substantial computing power. The red team’s experiment overcame that limitation by splitting the program across 15 televisions, effectively pooling modest compute resources to sustain functionality. The researchers said that if models become more capable on less powerful hardware — a trend they expect to continue — fully autonomous, self‑replicating malware could become feasible on ordinary devices.
Security experts have been receiving a steady stream of alarming and speculative AI risk advisories, complicating efforts to prioritize real threats. Anthropic’s report is notable because it documents a specific mechanism by which a malicious program could propagate without human intervention and sketches how near‑term improvements in model efficiency and device hardware might accelerate that risk. The company warned that such developments could make practical AI worms possible by the end of the year, calling for attention from companies and policymakers responsible for device security.
The report does not claim that a widespread attack of this type has occurred in the wild, nor does it provide evidence of active exploitation. Instead, it functions as an early warning intended to prompt further evaluation of smart‑device security, model deployment practices, and defenses that could limit unauthorized code execution or lateral movement across consumer hardware.