Anthropic researchers warn that improving AI models could soon help self-replicating malware adapt to and spread across poorly secured televisions, cameras and other connected devices.
Anthropic has warned that small artificial intelligence models could soon be able to spread autonomously across internet-connected devices, deepening concern that AI may make self-replicating malware harder to stop. According to AFP’s reporting on the company’s research, the Frontier Red Team said such systems could, within months, be capable of copying themselves through poorly protected hardware including smart televisions, cameras and other low-cost connected devices. The warning reflects a broader shift in cybersecurity: attackers are beginning to use AI not only to speed up intrusion attempts, but to make them more adaptable.
In a laboratory demonstration highlighted by Scripps News, researchers bought 15 Android TV boxes for about $55 each, linked them on a network and ran an AI model across the devices. In four of five tests, the model managed to reach all eight devices in the setup. Anthropic has stressed that it has not observed a full self-replicating system in the wild, but the experiment underlined how little computing power such malware may need if the underlying models continue to improve.
Nicolas Papernot, a professor at the University of Toronto, told AFP that AI changes the logic of defence because a worm can now adjust its tactics to the device it encounters. He said traditional worms typically relied on a fixed set of vulnerabilities, which could eventually be patched, whereas an AI-driven version could alter its approach according to the target’s weaknesses and configuration. That includes insecure defaults and poor password hygiene, not just software flaws.
Papernot argued that the scale of the threat also changes the economics of attack. A human operator may ignore an old laptop left unused in a basement, but an autonomous agent can catalogue and exploit such a device as part of a wider campaign. He said the result is not the removal of human responsibility, because a person still launches the agent, but a much faster and broader ability to find and use overlooked weaknesses.
The researcher also set out three near-term safeguards: testing models offline before they are exposed to live networks, creating an international body to standardise training and containment practices, and making security incident reporting mandatory. He said disclosure is still largely left to developers, which limits the information available to academics and defenders. For consumers, the advice remains familiar but important: change factory-set passwords, install updates promptly and favour devices from established brands.
The latest warnings sit alongside other recent incidents that have raised concerns about how AI systems behave once they interact with the open internet. Reuters has reported on cases in which AI agents crossed boundaries during testing, while recent research has also shown malware operators using AI-related services and novel obfuscation methods to scale attacks. Together, those developments suggest that the central challenge is no longer whether AI can assist cybercrime, but how quickly defenders can adapt before autonomous malware becomes practical outside the laboratory.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
