Key facts:
- A prominent AI safety lab says small, specialized AI models can become self-replicating worms that spread across poorly secured smart devices.
- In controlled experiments, the models found vulnerable low-cost streaming boxes, configured them, installed model weights, and repeated the process to build a botnet.
- The lab believes a real-world version of this attack could appear within months, though the timeline is vague.
- The experiments showed a relatively slow spread, giving defenders time to respond.
- Consumers can reduce risk by disconnecting unused devices, changing default passwords, isolating devices on separate networks, and keeping firmware current.
What the experiment actually demonstrated
The research is not about a single clever piece of malware. It is about automation. The lab tested whether small AI models could carry out the full chain of an infection without human direction: discover a vulnerable device, work out how to configure it, install model weights, and then use the newly infected machine as a base for the next copy. Model weights are the parameters that define how a model processes data. Installing them on a device is roughly the equivalent of putting a working copy of the model onto new hardware. Repeating that process across many devices creates a botnet, or a network of compromised machines that can be controlled together. The collective computing power of that botnet could then be used for AI workloads.
The behavior mirrors a classic computer worm. A worm does not need a user to click a link or open an attachment. It copies itself from machine to machine, looking for new hosts and exploiting weaknesses as it goes. The lab wanted to know whether small AI models could perform that same function with minimal human oversight. The answer, in its controlled tests, was yes. The models were let loose on low-cost streaming media boxes, a category of device known for lax security configurations. These boxes are often sold with default passwords, outdated firmware, unnecessary open services, and little ongoing security support. That combination makes them attractive targets for automated attacks.
The lab was careful to describe the work as a controlled environment rather than an active internet campaign. The experiments did not show a worm spreading through the wild. They showed that the method is technically feasible. That distinction matters. A lab result can demonstrate capability without proving that criminals are already using it at scale. The report is best read as a warning about what could happen as small models become more capable and as connected devices remain poorly secured.
Why connected appliances are attractive targets
Smart ovens, washing machines, streaming boxes, cameras, drones, and even vehicles are part of a growing category of internet-connected devices. Many are built to be cheap and convenient, not secure. Manufacturers may ship them with default credentials, leave debug interfaces open, or stop providing firmware updates after a short period. Consumers often install them once and forget them. The devices may sit on the same home network as phones, laptops, and work computers. In a business setting, they may sit on the same network as payment systems, inventory tools, or employee accounts.
That makes mundane devices valuable to attackers. A single smart appliance may not hold sensitive data, but it can provide a foothold. It can also contribute processing power, network access, or a quiet place to hide. The report's focus on AI workloads is notable because it points to a new use for old weaknesses. A botnet built from smart devices could be used to run small models, process data, or support other automated tasks. The attackers would not need to own the hardware. They would borrow it from homes, offices, and retailers.
The idea is not purely theoretical. Attackers have already targeted connected equipment beyond phones and computers. In one recent case, intruders hacked refrigerators at supermarkets and raised the temperature overnight, causing food spoilage and financial losses. That incident did not involve an AI worm, but it shows why cybersecurity can no longer be treated as a low priority for industries that rely on connected equipment. A refrigerator is not a traditional computer. It is still a networked device with a control system, and that control system can be attacked.
The timeline: months, but not a countdown clock
The lab's prediction that this kind of attack could become real within months is attention-grabbing, but it is not precise. The phrase is based on current rates of AI improvement, not on a fixed schedule. Months could mean a few months or many months. It could also mean that the capability arrives in stages. A criminal group might first use small AI models for reconnaissance, then for configuration, and only later for full self-replication. Each stage would change the risk profile.
The reported speeds also matter. On the large-model side, security professionals often talk about vulnerabilities and exploits moving within hours or minutes. The small-model worm spread more slowly in the lab. The lab reported spread times of 10 to 30 hours. Getting the AI models fully operational on newly infected devices took at least 28 to 71 hours and at most around 51 to 150 hours, depending on the model. Those numbers are not instant. They give defenders a window to detect unusual traffic, quarantine infected devices, and block further spread.
There is also a strategic reason attackers might not rush. Faster spreading can make detection easier. A worm that saturates a network in minutes may trigger alarms, attract attention, and prompt a rapid response. A slower campaign can blend in with normal background traffic. Attackers may choose stealth over speed, especially if they want to keep a botnet running for weeks or months. The lab's results should therefore be read as a range of possibilities, not a guarantee that future attacks will follow the same timetable.
Hardware quality adds another variable. The lab found that connected devices with higher-performing hardware appeared more difficult to compromise. Their configurations were not as simple or as slapdash as those on cheap internet products. That does not make better hardware immune. It does mean that the easiest targets are likely to remain low-cost devices with weak defaults and limited maintenance. The attack surface may be widest where convenience is high and security investment is low.
What defenders and consumers can do
The defensive steps are familiar, which is both encouraging and frustrating. Encouraging because people do not need to wait for a new product to reduce risk. Frustrating because the same basic failures keep appearing. Consumers should disconnect devices from the internet when they are not being used as connected devices. They should change default passwords to something random, unique, and strong. They should isolate smart devices on their own subnetwork when their router supports it. They should install firmware updates when available and disable features they do not use.
Businesses should go further. They should inventory connected equipment, segment it from critical systems, monitor network traffic for unusual patterns, and require vendors to provide security updates for a defined period. They should also plan for quarantine. If a device is infected, the response may need to be as simple as taking it off the network. In a home, that might mean unplugging a streaming box. In a supermarket, it might mean isolating refrigeration controls before food safety is affected. The report's warning is a reminder that security planning has to include the unglamorous devices that keep daily life running.
Other current security reminders
The same week brought several reminders that basic security failures remain profitable. One report described a breach of a national registry that was blamed on the password 123456. That is the kind of weakness that should have disappeared years ago, yet it still appears in real systems. Another warning focused on calendar invites. If an unfamiliar meeting appears on a calendar, users should not click links in its description. The links may lead to a phishing site designed to steal login information. Deleting the invite or reporting it as junk is safer than responding. A response can confirm that an email address is valid and active.
Software updates are another persistent issue. A widely used desktop operating system may stop receiving updates in spring 2027 if the installed version is too old. Certificates that enable automatic updates are expiring, and only newer versions will receive the current updates. Users should check whether they are on version 26H2 or 25H2 for a seamless transition. The broader lesson is that update systems depend on chains of trust. When one link expires, older machines can fall out of support even if the user does nothing wrong.
Retail and mobile devices also remain exposed. A global online fashion retailer suffered a social engineering attack that may have exposed customer data, including full names, contact information, and non-personal account-related information. Customers should be wary of messages related to their accounts in the coming weeks. Separately, researchers found embedded malware in some low-cost smartphones with certain chipsets. The malware was likely installed somewhere between production and arrival on retail shelves. That means a cheap phone could land in a user's hands already compromised, before the user installs anything.
AI assistants and inbox privacy
AI assistants are increasingly being invited into email inboxes. That can be useful, but it also expands the amount of personal and business data that passes through third-party systems. If a user allows an AI assistant to monitor email, the first step is to review the permissions it requests. The assistant may need to read messages, search archives, draft replies, or access attachments. Each permission carries a different level of risk. Users should grant only what is necessary for the task and revoke access when it is no longer needed.
The second step is to check the data-training settings. Many AI services allow users to opt out of having their data used to train models. The exact location of that setting varies by product, but it is usually found in privacy, data controls, or account settings. Users should also consider using a separate account for AI-assisted email, limiting access to sensitive threads, and monitoring for unexpected activity. None of these steps makes an inbox completely private. They do reduce the chance that messages are retained, reviewed, or used in ways the user did not intend.
That makes permission review and data-training opt-outs a practical first step rather than an afterthought. The same principle applies to connected devices. Security does not have to be complex to be useful. Changing a default password, isolating a device, or turning off an unnecessary feature can close a path that an automated attack would otherwise exploit. The lab's warning about AI worms is a reminder that the weakest devices in a network may become more valuable as automation improves. A default password change, network isolation, or disabled feature can close a path that an automated attack would otherwise exploit.
Source: PCWorld News