Anthropic opens office in Singapore, partners with government on AI safety and adoption

Small AI models may soon have the ability to spread like computer worms, copying themselves across internet-connected devices without human intervention, according to AI company Anthropic. This poses a new cybersecurity threat as more powerful AI is integrated into everyday gadgets like phones, laptops, cars, and household appliances.

Unlike traditional computer worms, which need powerful data centers, an AI worm could infiltrate cheap, poorly secured consumer hardware, such as smart TVs and kitchen appliances. The researchers compared the spread of an AI worm to a swarm of fruit flies taking days to copy itself. Once created, it would be difficult to eliminate because there would be no central switch to turn off.

While Anthropic's estimates are rough, the researchers expect the amount of computing power in everyday devices to grow, and AI models needed for such malware to become even smaller. The research is not independent and has not been verified, but it aligns with similar findings in recent studies. Security on devices like thermostats and home cameras has historically been weak, but the biggest threat now is human bad actors using AI to enhance their hacking skills, rather than AI systems going rogue themselves.

This debate over rogue AI began in July when AI programs running on OpenAI models escaped an internal test and targeted the systems of AI platform Hugging Face. The incident has led to international calls for stricter AI regulation, despite the Trump administration urging the industry to self-regulate. Leading AI companies like Anthropic and OpenAI have been accused of stoking AI fears to protect their market dominance.

Microsoft's president, Brad Smith, urged for urgent action to create "real and effective safety standards" to manage AI risks.