Anthropic has warned that artificial-intelligence-controlled worms could soon spread across smart gadgets. The statement continues a series of studies on how AI amplifies self-propagating malware.
A computer worm is malicious software that moves from one device to another on its own.
In June, Nicolas Papernot, an assistant professor of electrical and computer engineering at the University of Toronto, and a team built a prototype AI worm based on publicly available models. Papernot signed an open letter with recommendations for policymakers and told AFP why such attacks are harder to stop with AI.
According to him, many internet of things (IoT) devices - networked household appliances - have weak cyber defenses and easily become targets for attackers. Previously, a worm exploited a fixed set of vulnerabilities, and they could be patched to stop its spread. Now an attacker can create a worm that adapts its strategy to a specific device, and the vulnerabilities may differ. This significantly complicates interrupting the spread: all possible weak points have to be anticipated.
Some vulnerabilities are not coding errors but misconfigurations, such as the same password for different services or an incorrectly configured database. AI malware speeds up the exploitation of such vulnerabilities and increases the scale of attacks. If previously a coffee shop around the corner or a rarely used laptop might not be updated but did not attract an attacker's attention because of a lack of resources, now the autonomy of AI makes such devices interesting as a staging ground for attacks on more important targets. At the same time, Papernot stresses: an agent is launched by a human, and although after launch it can replicate, as shown in the work, the initiator remains human.
To reduce risks, the researcher proposes three short-term measures. First, change development and testing practices so that AI models are initially disconnected from the internet - this will allow their behavior to be better understood before critical infrastructure is connected. Second, create an international body to standardize the training and containment of AI models. Third, make reporting of security incidents mandatory: currently information comes only at the discretion of AI developers, while the academic community needs far more data to develop countermeasures faster.
Papernot acknowledges that the risk cannot be reduced to zero entirely, and even with fixes available, people do not always update devices in time. Despite decades of experts' calls, the world can still invest far more in the security of computer systems.