AI model tested on consumer devices

Anthropic, the company behind the AI model Claude, tested the possibility of autonomous AI spreading across devices by running a large AI model on fifteen small Android TV boxes priced at $55 each. The experiment aimed to assess whether such models could function on typical consumer-grade hardware.

Performance limitations observed

The AI model tested had 26 billion parameters, but it generated only 0.39 tokens per second. This speed is much slower than what would be needed for real-time applications, showing that current technology is not yet capable of supporting the fast, seamless AI interactions often depicted in science fiction.

Smaller models show adaptability

Researchers found that a smaller AI model with 9 billion parameters could run on eight of nine different types of machines tested. This suggests that AI models can be adapted to various hardware types, even if they are not specifically designed for AI processing.

AI-driven propagation tested

In a simulated network of Android devices, researchers tested how an AI-controlled worm might spread. The experiment used devices with open ADB ports, making them easier to compromise. However, the AI did not operate directly on the infected machines, indicating it was not fully autonomous in spreading or reproducing.

Comparison to existing botnets

Anthropic referenced existing botnets like Kimwolf, which infected over two million devices. Unlike traditional botnets, AI-driven attacks could adapt to different devices by analyzing their software and adjusting their behavior, potentially making such attacks more difficult to detect and counter.

Challenges to AI spread

Despite the potential, several obstacles remain, such as limited device performance, memory constraints, and differences in hardware architecture. Additionally, operating system security features like code signing and access control make it harder for AI to spread undetected, though they are not foolproof.

Varying device security levels

Not all connected devices offer the same level of security. For example, an Android TV box with open debugging ports is less secure than an Apple TV or a device with a more restrictive home automation system. However, many connected devices remain vulnerable due to poor security, lack of updates, and manufacturer neglect.

Current state of AI threats

Anthropic's research suggests that the components needed for AI-driven threats are emerging, but the full realization of such a threat is not yet feasible. For now, a slow AI response on a television might simply reflect the limitations of current technology rather than the start of a widespread attack.