A revolutionary threat: AI worm tailors attacks individually
A team of researchers from renowned institutions has developed a new generation of malware with the potential to fundamentally change the cybersecurity landscape. The adaptive AI worm uses large language models (LLMs) to develop tailor-made exploits for each target and to spread autonomously from device to device.
The international research team, made up of scientists from the University of Toronto, the Vector Institute, the University of Cambridge and ServiceNow Research, has published its findings in the paper "AI Agents Enable Adaptive Computer Worms" on arxiv.org. The results reveal a fundamental evolution of conventional computer worms.
How the adaptive AI worm works
Unlike well-known worms such as WannaCry, which target specific individual vulnerabilities, the new AI worm operates in a completely different way. It uses infected machines to run open large language models and thereby retains its decision-making capability while simultaneously increasing its reach for further attacks.
The decisive advantage of this approach lies in its cost structure: because the LLMs run on machines that have already been hijacked, the attackers incur no costs for computing power. Defenders, on the other hand, have to pay for AI-based countermeasures. In addition, this approach bypasses the central security checks and rate limits of commercial AI services.
Autonomous exploit development
Particularly remarkable is the worm's ability to develop exploits autonomously. The integrated LLMs analyse targets independently, adapt to new observations and develop attack logic in real time. This makes the worm independent of predefined exploit code or known proof-of-concept exploits.
Test results in a controlled environment
The researchers tested their AI worm in an isolated network of 33 devices from various categories. The test network comprised Linux servers, Windows environments and Internet of Things (IoT) devices. Known vulnerabilities that also occur in real corporate environments were deliberately left open.
Success rate and propagation
The test results across several seven-day runs showed impressive success rates:
- The worm correctly identified around one third of the vulnerabilities present
- On a quarter of the machines it managed to escalate its own privileges
- It propagated to around 20 percent of the devices
- The worm reached seven generations of self-replication
- In total it successfully attacked just under 75 percent of the network
- Replication succeeded on just under two thirds of the entire network
It was particularly notable that the worm independently developed working exploits for three vulnerabilities that only became known in 2026. This lay beyond the training cutoff of the LLMs used and demonstrates the capability to generate exploits autonomously from published information.
What this means for security technology
This development represents a paradigm shift in the threat landscape. While conventional worms can be stopped by applying security patches promptly, this strategy is far less effective against adaptive AI worms. The ability to adapt in real time and to develop exploits autonomously renders traditional defence strategies inadequate.
For operators of security technology, this means increased complexity in network monitoring and defence. Anomaly detection systems and behaviour-based security solutions are gaining in importance, since signature-based approaches are less effective against this adaptive threat.
Impact on critical infrastructure
The new class of threat could prove particularly problematic for critical infrastructure, where older systems with known vulnerabilities are often in use. The AI worm's ability to develop autonomous exploits even for newer vulnerabilities makes defence even more difficult.
Prevention strategies and protective measures
In view of this new class of threat, security managers must rethink their strategies. Multi-layered security concepts are gaining importance, since individual protective measures are less effective against adaptive attackers.
Network segmentation becomes a critical factor in limiting the speed of propagation. Monitoring systems should also be geared more strongly towards unusual network activity and anomalous system behaviour.
Continuous patch management
Although patches alone are no longer sufficient, prompt patch management remains essential. Reducing the attack surface by quickly closing known vulnerabilities makes it harder for the adaptive worm to gain a foothold and achieve its first successes.
Outlook for the future of cybersecurity
The researchers issue an urgent warning about the new class of threat: "We must prepare for autonomous generative adversaries: malware systems that spread without human involvement and that are not defined by fixed exploit code, but by the ability to analyse targets, adapt to observations and develop attack logic in real time."
This development marks the transition from reactive to proactive cybersecurity. According to the researchers, self-sustaining, AI-powered cyber threats are no longer mere theory but a real and immediate danger.
The need for new defence strategies
The security industry faces the challenge of developing defence mechanisms that can keep pace with the adaptability of AI-based attackers. This may require the use of dedicated AI systems for defence, which in turn brings new cost and complexity structures.
The research clearly shows that the cybersecurity landscape is facing fundamental change. Companies and institutions must reassess their security strategies and prepare for the challenges of adaptive, AI-powered threats.
