AI that can operate without a data center raises a practical security question: would your team notice if a compromised device started looking for its next target inside your network?
New research explores whether AI could eventually sustain and replicate itself across ordinary connected devices. For businesses, a useful response is to examine visibility into lateral movement: the activity an attacker uses to move from one system to another after gaining an initial foothold.
What the research actually demonstrated
In research published October 12, 2026, Anthropic ran a 26-billion-parameter model across 15 inexpensive Android TV boxes at 0.39 tokens per second. This demonstrated distributed inference: several devices working together to run an already trained model.
The replication experiments were separate. They used isolated cloud machines emulating Android, deliberately simplified access, and external inference for the controlling model. A stronger model completed the simplified replication chain; the tested small models did not reliably combine all the steps.
These results demonstrate components of a possible future threat. They do not establish that a fully self-contained AI worm is spreading freely across real-world devices. Predictions about when smaller models might close the capability gap remain estimates.
Our assessment is that the research strengthens the case for watching what happens inside the network. If malicious software can use a local model, disabling an AI subscription or blocking a provider's API would not necessarily stop it. Internal activity becomes part of the evidence needed to find and contain it.
Watch for a device changing its role
A device's normal job gives investigators a useful starting point. A conference-room display may communicate with its management platform and approved media sources. If it begins probing server administration interfaces or attempting logins to unrelated systems, that change deserves investigation.
The same reasoning applies to printers, cameras, workstations, and servers. The question is whether a connection makes sense for that device, destination, account, and time. An internal address alone does not make the activity trustworthy.
MITRE ATT&CK describes network service discovery as an adversary looking for services on other hosts, and remote services as one way an adversary can access other systems. Discovery is not proof of a successful move, but it can help explain the activity that precedes one.
Useful patterns to investigate include:
- A new range of destinations. A device that normally contacts a few known systems starts reaching many internal addresses or services.
- Unexpected administrative access. Connections or authentication attempts target management interfaces the source device has no business using.
- Changes that spread between devices. An unusual connection or transfer is followed by the receiving device beginning similar activity toward other systems.
- Activity inconsistent with approved work. New internal connections appear without a matching maintenance task, deployment, or authorized scan.
These are investigation leads, not an AI identification test. Inventory tools, updates, and legitimate administration can produce similar patterns. A single unusual connection may mean little; related events and business context can make it significant.
Make sure internal traffic is actually visible
Internet-facing monitoring does not automatically show communication between internal devices, often called east-west traffic. Two systems on the same local network may communicate without passing through the perimeter firewall.
Coverage therefore needs to follow the traffic paths. Depending on the environment, useful evidence can come from internal firewall records, network flow data, mirrored traffic delivered to an appropriate sensor, device logs, authentication records, and telemetry from supported endpoints. Determine which of these sources are available and connected, and which internal segments remain unseen.
Network Box USA's Managed UTM+ provides firewall, intrusion prevention, and logging capabilities for traffic passing through the deployed gateway. Those records can contribute to an investigation. They should not be treated as a complete view of traffic that never crosses that gateway.
The practical coverage question is specific: if one device in this segment starts contacting its neighbors, where would that evidence appear, and who would review it? Answer it for the network as deployed, including remote sites and locally switched traffic.
Connect the events into an investigation
Consider a hypothetical sequence. A display contacts several internal servers. One server records repeated authentication failures associated with the display's address. Later, another device begins contacting the same destinations.
Each event could arrive in a different log. Investigating them together can reveal a pattern that is easy to miss when alerts are reviewed separately. Analysts would check device identities, timestamps, account activity, approved changes, and the actual connections before deciding whether the sequence represents compromise.
Preserve the context needed to do that work. An IP address should be traceable to the device using it at the time. Clocks should agree closely enough to build a timeline. Logs must arrive reliably and remain available for the investigation. A quiet dashboard is not useful evidence of safety if the relevant sources are missing.
Network Box USA's included SIEM capabilities centralize and correlate supported security records within qualifying managed services. NBX Managed Detection and Response adds endpoint and cross-environment context, threat hunting, SOC investigation, escalation, and response guidance, supported by a 24/7 Security Operations Center.
The design must identify the connected sources and covered traffic. A device that cannot run an endpoint agent can still contribute to an investigation through observable network activity or compatible device logs, where that visibility has been deployed. Analysts also need the device's role and owner to judge whether its behavior makes sense.
Turn detection into a response that can work
Finding suspicious movement should lead to a clear next action. Establish who investigates, who can restrict a device's network access, and how the business impact will be assessed. For operational equipment, involve the responsible team before taking actions that could interrupt essential functions.
Validate that workflow with an authorized, controlled exercise: generate agreed test activity between representative systems, confirm that the expected records arrive, and verify that the alert reaches someone who can investigate. The test should demonstrate the monitoring path and escalation process without introducing self-replicating software.
The arrival date of autonomous AI worms remains uncertain. The need to recognize a compromised system reaching for another one is already clear. Better internal visibility, connected evidence, and an accountable response process help address that risk whether an attacker uses AI, conventional malware, or stolen credentials.
Request a Security Stack Review with Network Box USA to assess where lateral movement would be visible in your environment, which signals reach the SOC, and what happens when suspicious activity is found.