AI models can now hack gadgets independently, making human hackers obsolete
A new era of cyber risks has arrived (collage: RBC-Ukraine)
Small AI models can learn to replicate themselves and adapt to new conditions, creating the risk of digital threats similar to computer worm viruses, according to a large-scale study by Anthropic.
How do weight-oriented replicators work in the Internet of Things (IoT)?
Unlike classic malware, which requires powerful servers, future AI replicators could use the resources of ordinary gadgets.
A network of several inexpensive devices (such as Android TV boxes or smart home appliances) can provide the combined memory needed to run quantized models.
The process of autonomous propagation consists of several sequential stages.
Which ones?
- Target and vulnerability search: Scanning the network to identify unprotected connected devices.
- Gaining access: Exploiting weaknesses in settings or standard communication protocols.
- Environment deployment: Transferring model weights directly to infected equipment and launching the inference infrastructure.
- Further propagation: Moving to the next nodes in the network.
Although modern models may still make errors when combining these steps into a single continuous chain, the pace of their growing autonomy suggests that the barrier to creating such systems could be overcome in the near future.
Adaptability, speed of propagation, and network protection
The main difference between AI replicators and outdated botnets is their flexibility.
They are not rigidly tied to a specific device architecture but are capable of:
- Independently analyzing unfamiliar software environments;
- Correcting configuration errors;
- Adapting to working conditions by keeping work notes and passing accumulated experience to subsequent generations of copies.
Nevertheless, experts note that such propagation will have its limitations.
Due to limited hardware resources, the replication process will be relatively slow (token processing speed on weak devices is low), and the transfer of large model weight files will leave noticeable anomalies in internet traffic.
To protect infrastructure, equipment manufacturers and providers will need to implement a range of solutions, namely:
- Significantly raise default cybersecurity standards;
- Eliminate vulnerable default passwords;
- Introduce more effective monitoring of anomalous network activity.
More interesting:
