Artificial intelligence is changing the cybersecurity landscape on both sides of the attack.
Security teams are using AI to analyze threats, accelerate vulnerability research, improve detection, and streamline security operations. At the same time, researchers are demonstrating how artificial intelligence could give attackers new ways to automate and adapt malicious activity.
One of the most concerning developments is the potential for AI-powered computer worms.
Researchers at the University of Toronto have demonstrated a proof-of-concept AI worm capable of adapting its attack strategy to different online devices in a controlled research environment. The research highlights a potential evolution in malware: instead of relying entirely on predefined instructions, an AI-powered worm could use an AI model to analyze its environment and generate attack strategies tailored to individual targets.
For security teams, the issue is not simply whether AI can help attackers write malware. It is whether AI can make attacks more adaptable, scalable, and difficult to predict.
A computer worm is a type of malware designed to replicate itself and spread from one computer or device to another.
Unlike a traditional virus, a worm can often spread without requiring a user to manually execute an infected file. Historically, worms have exploited vulnerabilities in software, operating systems, network services, and other technologies to move from one system to another.
Traditional worms typically rely on predefined instructions.
A threat actor may program a worm to identify a specific vulnerability, exploit it, establish access, and search for additional systems. If organizations patch the vulnerability or otherwise disrupt the attack path, the worm may become significantly less effective.
An AI-powered worm introduces another possibility. Instead of following one rigid sequence of instructions, an AI-enabled system could potentially analyze information about its target, identify potential attack paths, generate code or commands, and modify its approach based on what it encounters.
The University of Toronto research demonstrates this concept in a controlled environment, showing how an AI worm could generate tailored attacks for different targets.
The significance is not that AI has suddenly created unstoppable malware. It is that the technology could lower some of the barriers involved in developing adaptive attacks.
Threat actors can automate reconnaissance, vulnerability exploitation, credential theft, privilege escalation, and lateral movement. They can also use malware frameworks that contain numerous tools and capabilities.
AI adds another potential layer of adaptability. Consider the difference between these two approaches:
A traditional worm might effectively say, "If I find vulnerability X, execute technique Y." An AI-enabled worm could potentially operate more like, "I have encountered this environment. What attack paths are available to me, and which one is most likely to work?"
If a specific vulnerability is unavailable, an adaptive system could potentially search for another route. If an attack attempt fails, it could potentially generate another strategy. If two targets have different configurations, the system could potentially tailor its behavior to each. This could make defensive strategies based solely on known indicators less reliable.
The goal of AI-powered malware is not necessarily to invent completely new hacking techniques. In many cases, its value to an attacker could simply come from automating and adapting techniques that already exist.
Cybersecurity defenses often depend on identifying suspicious or malicious behavior.
Security teams use endpoint detection, network monitoring, vulnerability management, identity controls, intrusion detection, threat intelligence, and other technologies to identify and stop attacks.
However, adaptive malware creates a challenge because the same underlying attack could potentially look different across environments.
A threat actor may use different techniques depending on:
Operating system
Available software
Network configuration
User privileges
Security controls
Exposed services
Available credentials
Cloud infrastructure
Connected devices
The more a threat actor can adjust to those variables, the less useful a one-size-fits-all defensive assumption becomes.
This is one reason organizations should increasingly focus on behavior and attack surface, rather than relying exclusively on signatures or known malware characteristics.
One of the biggest potential advantages AI offers attackers is scale.
Cyberattacks require resources. Skilled operators need to conduct reconnaissance, analyze environments, research vulnerabilities, write or modify code, and determine what to do next.
AI could automate portions of that work.
An AI-enabled attack system could potentially assist with:
Reconnaissance
Target identification
Vulnerability analysis
Attack-path selection
Code generation
Exploitation
Post-exploitation analysis
Lateral movement
Target prioritization
Not every step can or will be fully autonomous. However, even partial automation can have significant consequences when applied at scale.
The concept of AI-enabled worms is not completely new. In fact, researchers have previously explored whether malicious prompts could propagate between generative AI systems. The emergence of increasingly autonomous AI agents has expanded the conversation further.
Today, AI systems can interact with applications, APIs, files, databases, and other digital resources. That creates a fundamentally different security environment.
An AI system is no longer necessarily just a chatbot responding to a user. Depending on how it is configured, an AI agent may be able to take actions on a user's behalf.
Organizations increasingly rely on interconnected technologies. Cloud environments, SaaS applications, APIs, remote access systems, identity platforms, endpoints, third-party services, and AI applications can all form part of an organization's attack surface.
A vulnerability in one system may therefore provide an attacker with a path toward another. AI does not eliminate these traditional security risks.
Instead, it could potentially make it easier for attackers to identify and exploit relationships between them. This is why vulnerability scanning alone is not enough.
A scanner might identify an individual vulnerability. A penetration test can help determine whether that vulnerability can actually be exploited, whether it can be combined with other weaknesses, and what an attacker could ultimately reach.
As attackers become more adaptive, organizations need to understand whether their defenses actually withstand realistic attack behavior.
A penetration tester may discover that an apparently low-risk vulnerability can be combined with a weak authentication control, excessive permissions, or an exposed administrative interface to produce a much more serious compromise. That context is critical.
Packetlabs uses a human-led penetration testing approach supported by targeted automation and AI. Its infrastructure penetration testing methodology emphasizes experienced testers validating exploitability, connecting weaknesses across identity, cloud, and internal environments, and focusing on real-world impact.
The rise of AI-powered cyberattacks does not mean human cybersecurity professionals are becoming irrelevant.
As AI becomes more capable, human judgment becomes increasingly important for interpreting what the technology discovers.
AI can help process information quickly. It can identify patterns, generate potential approaches, summarize technical data, and automate repetitive work. But a penetration tester still needs to determine:
Is this vulnerability actually exploitable?
Can the weaknesses be chained together?
What access does the attacker gain?
Can the attacker move laterally?
What sensitive systems are reachable?
What controls prevent further compromise?
What is the business impact?
What remediation should be prioritized?
These questions require technical knowledge and contextual judgment.
Packetlabs' current approach reflects this distinction. The company describes its penetration testing as human-led, with AI used under direct tester oversight. Packetlabs also states that client data does not touch public AI services or train models, with those commitments documented in its statements of work.
It is important to distinguish AI-enabled penetration testing from automated vulnerability scanning.
A vulnerability scanner may identify a known vulnerability based on a signature or technical indicator. AI can make scanning and analysis more efficient, but finding a vulnerability is only one part of understanding security risk.
Penetration testing goes further. A qualified tester attempts to validate whether weaknesses can be exploited and determine how they interact with the broader environment.
For example, an attacker might combine:
A compromised credential
An overly permissive identity
An exposed internal service
A vulnerable application
Weak network segmentation
Individually, these issues may not appear catastrophic. Together, they could create a path to sensitive systems.
That is why organizations evaluating AI-enabled security testing should ask whether the provider is actually validating attack paths or simply automating vulnerability discovery.
There is another side to the AI security equation. Organizations are not only defending against attackers who use AI. They are increasingly deploying AI themselves.
AI assistants, large language models, autonomous agents, AI-enabled applications, and model APIs can introduce new vulnerabilities. Potential risks include:
Prompt injection
Jailbreaking
Sensitive data exposure
Insecure integrations
Excessive permissions
Model manipulation
Insecure API configurations
Unauthorized access to connected systems
Packetlabs' AI and LLM penetration testing services specifically address these types of risks. Its testing includes areas such as prompt injection, data leakage, model abuse, and API exposure.
Organizations do not need to wait for an AI-powered worm to become a widespread incident before taking action.
The same foundational security practices that protect against conventional attacks remain important.
AI does not make traditional vulnerabilities irrelevant.
Organizations should maintain accurate asset inventories, identify exposed systems, prioritize vulnerabilities based on risk, and deploy security updates promptly.
An adaptive attacker can potentially make greater use of weaknesses that organizations leave unaddressed.
Organizations should understand what is exposed to the internet and why.
Unused services, unnecessary accounts, outdated systems, exposed administrative interfaces, and excessive permissions can all create opportunities for attackers.
Reducing unnecessary exposure makes adaptive attacks harder to execute.
Identity is increasingly central to cybersecurity.
Organizations should review privileged accounts, service accounts, cloud permissions, third-party access, authentication controls, and dormant accounts.
Multifactor authentication should be deployed wherever appropriate, particularly for privileged access.
Network segmentation can limit how far an attacker can move after compromising an endpoint or account.
This becomes particularly important when considering malware designed to spread automatically.
Security teams should look beyond known malware signatures.
Unexpected processes, unusual authentication activity, suspicious network connections, abnormal privilege use, and unexpected access to sensitive resources can all provide important signals.
Security controls should be tested, not simply deployed.
Penetration testing can help determine whether controls prevent realistic attack paths and whether attackers can bypass them.
Packetlabs describes its broader penetration testing services as focused on identifying and prioritizing real attack paths and validating defenses against evolving threats.
As AI changes the threat landscape, security leaders should also change how they evaluate penetration testing providers.
Instead of asking only whether a provider uses AI, ask: How is AI being used? Is it supporting reconnaissance and analysis, or is it being given unrestricted control over sensitive systems?
Who validates the results? AI-generated findings should not automatically become confirmed vulnerabilities.
How is customer data handled? Organizations should understand what information is provided to AI systems and where it is processed.
Is the testing human-led? Human expertise remains critical for validating exploitability and understanding business impact.
Can the provider test realistic attack paths? The ability to identify vulnerabilities is different from demonstrating how an attacker could chain them together.
Does the provider have independent assurance? Security leaders should consider relevant certifications and attestations when evaluating vendors.
Packetlabs is CREST-accredited and SOC 2 Type II attested, and in 2026 became one of the first organizations to achieve CREST's AI-Enabled Penetration Testing accreditation.
AI-powered computer worms represent a broader shift in the cybersecurity landscape.
The technology demonstrated by researchers does not mean organizations are facing an imminent wave of autonomous malware. But it does show that AI can potentially be used to make malicious software more adaptive. That matters because cybersecurity is already moving away from a simple perimeter-defense model.
Organizations have increasingly complex environments, distributed identities, cloud infrastructure, third-party integrations, and AI-powered applications.
Attackers are looking for connections between those systems. Security teams need to do the same. The most effective security strategies will increasingly focus on understanding how individual weaknesses combine into real attack paths.
The emergence of AI-powered computer worms is a reminder that cyber threats are becoming increasingly dynamic.
Traditional malware can already automate attacks at significant scale. AI introduces the potential for greater adaptability, allowing malicious systems to analyze their environments and adjust their behavior.
Organizations should not respond by abandoning traditional cybersecurity practices. They should strengthen them.
Vulnerability management, identity security, network segmentation, endpoint monitoring, security testing, and incident response remain fundamental. But organizations also need to validate whether those controls hold up against realistic, adaptive attack scenarios. That is where human-led penetration testing continues to matter.
AI can help security professionals work faster and analyze more information. It cannot replace the judgment required to determine what an attacker could actually accomplish inside a real environment.
For organizations navigating an increasingly AI-driven threat landscape, the goal should not be to choose between AI and human expertise.
Ready to pressure-test your defenses against evolving attack paths? Talk to Packetlabs about your penetration testing requirements.
An AI-powered computer worm is a type of malware that could use artificial intelligence to analyze target environments, generate or adapt attack strategies, and potentially spread between systems. Current research has demonstrated the concept in controlled environments.
There is no indication from the University of Toronto research that an autonomous AI worm is currently spreading across the internet. The research was conducted in a controlled digital environment as a proof of concept.
Traditional worms generally rely on predefined instructions and exploits. An AI-enabled worm could potentially adapt its behavior based on the environment it encounters, potentially allowing it to select or generate different attack strategies.
Penetration testing cannot guarantee protection against every future threat. However, it can help organizations identify vulnerabilities, validate security controls, and understand how an attacker could move through their environment.
Yes. Packetlabs states that AI is used under direct tester oversight as part of its human-led penetration testing approach. The company also states that client data does not touch public AI services or train models.
Yes. Packetlabs provides AI and LLM penetration testing that assesses risks including prompt injection, data leakage, model abuse, and API exposure.
AI-enabled penetration testing refers to the use of AI by security professionals to support the penetration testing process. AI/LLM penetration testing refers to testing an organization's AI systems and applications for vulnerabilities.
Human testers provide the technical judgment needed to validate vulnerabilities, chain weaknesses together, assess business impact, and determine whether an attack path is realistically exploitable.
Organizations should maintain strong vulnerability management, reduce unnecessary attack surface, secure identities, segment critical systems, monitor for abnormal behavior, and regularly validate security controls through realistic penetration testing.
Our Penetration Security Testing methodology is derived from the SANS Pentest Methodology, the MITRE ATT&CK framework, and the NIST SP800-115 to uncover security gaps.
Download our Pentest Sourcing Guide to learn everything you need to know to successfully plan, scope, and execute your penetration testing projects.
AI watermarking is rapidly becoming a cybersecurity and content-trust priority. Learn how Anthropic, Google and C2PA are making AI content traceable and the limitations of watermarking.
September 21, 2026 - Blog
Ontario courts were affected by a cyberattack involving Thomson Reuters' C-Track case management platform. Learn what happened, what data may be exposed, and what organizations can learn from the breach.
September 11, 2026 - Blog
Explore five of the biggest cyber attacks of 2026, including Cisco SD-WAN, Ivanti EPMM, Stryker, and ShinyHunters attacks, and learn what DNS threat intelligence reveals.
September 04, 2026 - Blog