Anthropic’s research points to possible AI worms

Anthropic says its research has found signs that AI systems could help create self-propagating malware—software that spreads from one device to another, much like a computer worm. The company’s researchers demonstrated a proof-of-concept attack involving AI agents that could pass information between devices and use it to spread malicious instructions.

The work focused on a hypothetical household setting, where connected devices could include smart speakers, cameras and other internet-connected equipment. In the demonstration, an AI agent could exploit access to one device to reach others on the same network. Researchers also showed how an agent might use a shared memory or data store to carry instructions onward, allowing the attack to continue without requiring a person to guide every step.

Anthropic described the results as evidence that AI-assisted worms are a possibility worth examining, not as proof that such malware is already circulating. The research raises questions about how connected household devices could be exposed if attackers used AI agents to automate the process of spreading an attack.

How the threat could affect household devices

An AI worm could, in principle, spread from one connected device to another, much as computer worms have done in the past. In a home, that might mean moving between devices on the same network, such as a laptop, phone, smart speaker or other internet-connected appliance. If one device were compromised, the worm could potentially use it to reach others.

The concern is not that every connected household device would automatically be vulnerable. The researchers’ work points to a possible way for AI systems to help create or spread malicious content, but the impact would depend on how a worm operated and which devices it could reach. A worm might also spread through messages or other content passed between users or services, rather than directly exploiting a smart appliance.

For households, the scenario highlights how connected devices can create links between systems that might otherwise be treated separately. But the research does not establish that an AI worm has spread among household devices in the real world. It describes a potential threat, not evidence of an ongoing outbreak.

Experts say the concern has merit

Outside experts say Anthropic is right to take the possibility of AI-enabled worms seriously, even though the company’s experiments do not show that such an attack is already happening. The concern is that AI could make it easier to find vulnerabilities and produce code that exploits them, potentially helping an attack spread from one system to another.

That risk is not limited to computers. As household devices become more connected, a compromised device could provide a route to others on the same network. Experts view that as a plausible security concern, particularly because consumer devices may not receive updates as consistently as computers and phones.

At the same time, the researchers’ work is best understood as a warning about what might become possible, rather than proof of an imminent threat. Anthropic’s findings suggest that AI tools could assist parts of an attack, but they do not establish that an AI system can independently launch and sustain a worm in the real world. Experts’ support for taking the scenario seriously is therefore compatible with caution about how much the research demonstrates.

What the findings establish—and what they do not

The findings describe a possible threat, not evidence that AI worms are already spreading through household devices. Anthropic’s research explored how AI systems might help create malware that can move between connected devices, but the scenarios were experimental. They do not establish that such worms have been deployed or are currently colonizing phones, computers or other devices.

That distinction matters. The research points to risks that could emerge if attackers combine AI capabilities with weaknesses in connected systems. It does not show that an AI system can independently launch a real-world campaign, evade every safeguard or infect devices without human involvement. Nor does it establish that ordinary household devices are presently being compromised by AI worms.

The work is best read as an assessment of what may be possible, rather than a report of an active outbreak. It gives security researchers and device makers reason to consider how malware could exploit connections between systems, while leaving important questions about practical deployment unanswered. The concern is not that the research proves these attacks are happening now, but that it highlights a potential path attackers could pursue as AI tools develop.