Why This Matters
- If self-replicating AI worms become viable, autonomous malware could spread across the wider internet of things, including home appliances and other connected devices that are often less protected than computers and phones.
- The warning comes from a leading AI developer, signalling that the companies building agentic AI are treating autonomous spread as a realistic near-term risk, not a distant hypothetical.
- For users, the practical question is how quickly platforms and device makers introduce safeguards around AI systems that can act on their own.
Background
Anthropic is a leading artificial intelligence company known for its Claude line of AI models. Malicious software that spreads between devices is a long-established threat, but AI changes the picture: AI agents can be given the ability to read content, browse the web and operate software, which expands the ways automated attacks could unfold. Security researchers have demonstrated proof-of-concept worms that spread between AI assistants, and prompt injection, a technique where hidden instructions embedded in text steer a model into unintended actions, is a recognised weakness of the technology.
Key Perspectives
Anthropic: The company is publicly warning that self-replicating AI worms are months away, framing autonomous spread across user devices as an urgent security challenge that needs attention before it arrives.
Security researchers: Experts who study AI security have long argued that giving models access to external tools and the internet expands the attack surface, and that safeguards such as sandboxing and permission limits will determine how serious the threat becomes.
Critics and Skeptics: Skeptics caution that predictions about near-term AI capabilities are inherently uncertain, and that the real-world risk depends on AI systems being granted the permissions needed to act across different devices, which users and platform operators still control.
What to Watch
- Whether researchers demonstrate a self-replicating AI worm operating outside a tightly controlled lab environment.
- How device makers and AI platforms respond with restrictions on autonomous actions, such as requiring confirmation before an AI system interacts with other devices.
- Further public assessments from Anthropic or other AI companies that confirm, refine or push back on the months-long timeline.