MUMBAI: The next computer worm may not need a hacker pulling the strings, it could write its own playbook as it goes. A new study from researchers at the University of Toronto has sounded a fresh alarm for cybersecurity experts, demonstrating how artificial intelligence could enable a new generation of self-spreading malware capable of adapting, learning and attacking networks without human intervention.
Detailed in a research paper titled AI Agents Enable Adaptive Computer Worms, the study outlines a prototype AI-driven worm that can autonomously move between devices, analyse vulnerabilities and develop customised attack strategies in real time.
The findings point to a potentially significant shift in cyber threats. Unlike traditional computer worms, which typically rely on exploiting a single known flaw, the AI-powered version continuously assesses each target, identifies weaknesses and adjusts its tactics accordingly. In practice, that means defenders may no longer be able to stop an attack simply by patching one vulnerability.
Researchers described the system as an example of how advances in artificial intelligence are beginning to intersect with offensive cybersecurity capabilities. Rather than following a fixed set of instructions, the worm uses AI models to make decisions independently, enabling it to adapt as conditions change across a network.
What makes the development particularly notable is its reliance on open-source large language models. By using publicly available AI systems instead of commercial platforms, the prototype operates independently of many safeguards that technology companies typically impose, such as usage restrictions, rate limits and monitoring mechanisms.
To evaluate its effectiveness, researchers tested the worm in a simulated corporate environment containing a mix of Linux, Windows and Internet of Things (IoT) devices. The results were striking.
According to the study, the malware successfully compromised nearly 75 per cent of the network within a week, establishing a persistent presence across multiple systems and continuing to spread without direct human control.
Once inside a device, the worm was able to utilise the machine’s computing resources to support further operations and attacks. Researchers noted that this dramatically reduces the ongoing cost of propagation for attackers, potentially allowing malware campaigns to scale more efficiently than conventional approaches.
The implications extend well beyond corporate networks.
Because the AI system can tailor its attack methods to different environments, virtually any internet-connected device from laptops and servers to smart cameras, home automation systems and connected appliances could become a potential target.
Cybersecurity specialists have long worried about the combination of artificial intelligence and malware. This research offers a glimpse of what that future could look like, attacks that are not only automated but capable of evolving as they spread.
The study also highlights a challenge facing defenders. Traditional cybersecurity strategies often revolve around identifying a vulnerability, deploying a patch and blocking further exploitation. An adaptive AI worm changes that equation by constantly searching for alternative pathways, reducing reliance on any single attack method.
Recognising the potential risks, the University of Toronto team conducted the research in a secure, isolated environment and withheld key technical details that could facilitate misuse. The researchers stressed that the objective was not to provide a blueprint for attackers but to alert governments, businesses and cybersecurity professionals to an emerging threat landscape.
As artificial intelligence becomes more powerful and widely accessible, the race between attackers and defenders appears set to enter a new phase. The concern is no longer simply whether malware can spread automatically, it is whether future malware can think its way around the obstacles designed to stop it.
For the cybersecurity industry, that possibility may be the most unsettling vulnerability of all.
