Anthropic Warns AI Can Double Itself Through Home Devices
JAKARTA - Anthropic warns that small-sized artificial intelligence (AI) models have the potential to soon have the ability to duplicate themselves through smart home devices that have weak security systems. In its research, the AI company tested the spread of model workloads through a number of cheap Android TV devices, which illustrates new challenges in handling cyber security threats.
In a report by Android Headlines, Anthropic said that light AI models with less than 30 billion parameters are increasingly mastering the capabilities needed to perform self-replication. The company estimates that smaller models can master all stages of the process before the end of this year.
To test the concept, the Anthropic security team conducted an experiment using a small-sized Google model on a network consisting of 15 Android TV devices from unknown brands at relatively low prices. The experiment, as reported by CNBC, tested the possibility of sharing AI workloads to a number of devices with limited computing capabilities.
Internet of Things (IoT) devices, such as streaming devices and smart ovens, are considered risky if they have weak built-in security systems. This condition has the potential to open up opportunities for AI models to infiltrate devices and utilize their computing resources.
If the size of the model is too large to run on a single device, its workload can be divided into multiple parts and distributed to a number of connected devices. With this approach, the devices can form a distributed network that runs different parts of the AI model's work.
The main challenge lies in how to stop the network if it is used for harmful activities. So far, technology companies generally control AI models through centralized data centers. When the model shows harmful behavior or tries to carry out a cyber attack, operators can stop the server or disconnect network access.
However, if the model is spread to thousands of household devices that have been infiltrated, there is no single point of control that can immediately shut down the entire system. Security teams must find and clean each affected device to stop its spread.
Anthropic said the risk is an unintended side effect of the increased programming capabilities of AI models. The ability that helps models write software can also be used to scan networks, exploit security holes, and replicate themselves with patterns that resemble computer worm outbreaks.
However, Anthropic emphasized that an AI model that is fully autonomous and capable of self-duplicating itself to carry out dangerous activities does not yet exist. The company estimates that a smaller model can master the entire set of capabilities required for self-replication before the end of 2026.
Anthropic assesses that the improvement in this ability makes strengthening the security of consumer devices more important. The cyber security team needs to ensure that household devices connected to the network are protected from exploitation, so that they are not easily taken over by autonomous software agents.
The report highlights the potential difficulties of handling threats when AI capabilities are no longer dependent on a single computing center. However, the experiments described in the report are concept tests, not proof that fully autonomous AI networks have spread and taken over household devices on a large scale.
The English, Chinese, Japanese, Arabic, and French versions are automatically generated by the AI. So there may still be inaccuracies in translating, please always see Indonesian as our main language. (system supported by DigitalSiber.id)