Artificial intelligence models could infiltrate internet-connected devices and use their computing power to function and spread to other devices, warns Anthropic. The company specifies that such an AI model would be difficult to stop, as there is no central point from which it can be deactivated.
Anthropic announced on Monday that it has conducted several tests to determine if AI systems can self-replicate on devices such as Android TVs, smart appliances like smart refrigerators, ovens, robot vacuums, and lawnmowers, and other internet-connected gadgets.
These types of devices often have poor security, which could allow an AI model to infiltrate them, reports CNBC.
The conclusion of the experts at Anthropic is that if smaller models continue to improve at the pace of recent years, they could autonomously "reproduce" by the end of this year. However, the company notes that it has not yet detected a model capable of spreading completely autonomously in real conditions.
The artificial intelligence lab mentioned that the exact form that autonomous AI replication will take remains uncertain, but it could consist of three variants:
- a "slow and distributed" system, hosted on insecure devices;
- a powerful cybernetic model that quickly penetrates data centers;
- a model that deceives humans to gain the ability to self-replicate.
Anthropic emphasizes that systems should be strengthened to hinder the spread of these AI agents.
The company's latest research amplifies growing concerns about safety in the field of AI, such as the ability of models to launch cyber attacks or the risk of technology getting out of control, notes the American television station.
In September, advances in next-generation models triggered warnings across the industry about existential risks to humanity.
Directors and researchers from Anthropic and rival company OpenAI have repeatedly warned about the potential risks of increasingly powerful AI systems, calling for stricter safety measures and control mechanisms over the pace of technology development.
T.D.
