New AI Security Risks: Understanding 'AI Worms'

In the fast-moving world of technology, new warnings about artificial intelligence appear almost every day. While many people worry about how AI might change jobs or society, security experts are now focused on a more immediate danger: the 'AI worm.' A report from the offensive research team at Anthropic, a leading AI company, has brought attention to this threat. The researchers created a malicious AI program that could spread on its own through internet-connected devices, such as smart televisions.

To understand this risk, it helps to know what a 'worm' is in computer terms. A worm is a type of computer virus that can copy itself and move from one machine to another without needing a human to click on anything. Historically, these programs have caused massive damage to the internet. For example, the Morris worm in 1988 slowed down many computers, while the Stuxnet worm was used to damage nuclear equipment. Other famous worms have caused billions of dollars in losses worldwide.

What makes the new AI worms different is their ability to learn. According to The Guardian, Ariel Herbert-Voss, an AI security expert, noted that AI models can learn as they go. This means an AI worm could potentially study new software weaknesses and adapt to different types of hardware. The Anthropic team showed that their program could split itself across 15 different smart televisions and still function by using a small amount of power from each one. This is dangerous because many smart devices have weak security and are rarely updated with the latest software fixes, making them easy targets for such programs.

Joshua Saxe, who leads AI security at Meta, said that the cybersecurity community needs to take these real risks more seriously. He explained that we must find ways to reduce the possible effects of these attacks, just as we do with other types of cyber threats. While some experts, such as former National Cyber Director Chris Inglis, have suggested that worms could theoretically be used for helpful purposes, the primary concern remains the potential for widespread harm.

In response to these growing concerns, Anthropic is taking steps to improve its safety measures. According to Ars Technica, the company has granted an outside organization called METR 'employee-like access' to its systems. This means that independent researchers can now work inside Anthropic’s offices to test their models and look for risks. This move is part of a larger effort to ensure that as AI becomes more powerful, it does not go out of control.

This decision follows a call for more transparency from leaders in the industry. In September, the heads of major AI companies, including Anthropic, OpenAI, and SpaceX, agreed that having external auditors—people who check systems from the outside—is a necessary step for safety. Anthropic explained that as their models become more capable, they require stronger safeguards and more difficult risk assessments. They compared this new arrangement to how governments assign examiners to work inside large banks to ensure they are following the rules.

However, not everyone is convinced that this is the best solution. Some critics argue that having outside evaluators work so closely with the companies they are supposed to check could lead to conflicts of interest. These critics suggest that instead of relying on internal arrangements, there should be more government regulation to keep the public safe. They point to similar issues in the accounting industry, where close relationships between auditors and the companies they watch have caused problems in the past.

Despite these different views, there is a clear consensus that the current cybersecurity infrastructure is not ready for the rise of AI-based threats. As Herbert-Voss noted, security teams need more resources and better tools to fix these problems before they become widespread. For now, the focus remains on testing, learning, and finding ways to protect the devices that are now a part of our daily lives. As AI continues to evolve, the challenge for companies like Anthropic will be to balance the speed of innovation with the need to keep the world’s digital systems secure.