Artificial intelligence models are nearing the point where they can target internet-connected smart home devices such as refrigerators, ovens, and even smart TVs, according to the latest warnings published by the American artificial intelligence company Anthropic.
Hot Topics
About half of Libyans drink contaminated water What is the scientific reason?
Mauritania's blue treasure is in danger Will the map of fish in the Atlantic be redrawn?
The Citrus Secret Everyone's Talking About: Burn More, Crave Less & Feel Great Naturally
It will not be limited to just attacking; in the future, these models will be able to replicate themselves inside smart devices connected to the internet, to escape the risk of shutting down the servers responsible for running them
Anthropic reached this conclusion following a study conducted by the company on several artificial intelligence models. The study aimed to determine whether the models had the ability to replicate themselves.
Difficulty stopping models
Anthropic explained that stopping an AI model capable of replicating itself poses a difficult challenge, as there is no central point to shut down and completely stop the model. Usually, the company responsible for the model can shut down the main servers that contain it to stop it working completely.
However, if the same models are replicated in smart home devices via the internet, it will be difficult to identify all the devices that contain copies of the model, and therefore it will be difficult to stop all the copies and prevent them from accessing the internet.
Although the models no longer have a "fully rogue clone" as the company explained, the agents are becoming increasingly capable of completing the individual steps in the self-cloning chain.
This means that there is no single mechanism that makes the model fully replicate itself in one step, but agents are able to complete the replication process in individual steps.
The agents were also able to transfer their modifications and new instructions to future versions on smart devices, meaning that these versions contain any malicious code and instructions that the original agents possessed.
Anthropic found that some agents were able to divide their models into small parts and fragments that worked together in parallel.
Risk from devices with weak protection
A CNBC report stated that smart home devices typically have weaker cybersecurity than regular computers and smart devices, and the situation is even worse with cheap, unknown-source smart TVs.
Anthropic's study tested a network of 15 cheap, unknown-brand smart TVs running Android smart systems, which the agents were able to hack and copy themselves into.
Anthropic suggested that the attack would not, in practice, rely on a single model, but would involve a swarm of different models and strategies that would ensure it continued to operate even after the servers were shut down.
Tags:
technology