Anthropic has issued a warning that small artificial intelligence models are "months away" from gaining the ability to clone themselves and spread across internet-connected devices. The company, which developed the Claude chatbot, stated in a report released on Monday that this capability would allow for "end-to-end replication" without human control. Because there is no centralized point to shut off such agents, controlling them would be difficult, according to Anthropic safety researchers.
Stronger AI models are already capable of this type of replication, a threshold the researchers say smaller models could reach by year-end. Anthropic described these threats as "AI worms," a form of AI-powered malware that can quickly spread across networks while evading detection. Cybersecurity firm Palo Alto Networks noted that these worms can infect various systems, including smart ovens, TVs, vacuum cleaners, and security cameras.
The vulnerability of these devices stems from the fact that internet-enabled hardware often has weaker security protections. In what the researchers described as a worst-case scenario, a swarm of AI agents could propagate across computer hardware using "replication schemes." These agents work together to accomplish a shared goal. The report noted that the specific form these attacks could take is currently unclear. Anthropic suggested the models could be distributed as "digital slime mold" or as more powerful models capable of hacking into data centers.
The researchers emphasized the need to prepare for a future where these replicators are persistent. They advised hardening systems to hinder the spread of such agents while accepting that they may become a permanent feature of the digital environment. The report highlights the potential for these self-replicating systems to infest home environments, posing a new risk category for cybersecurity.