Self-replicating AI possible within months, Anthropic warns

AI models are getting closer to replicating themselves

The company’s latest tests examined whether AI systems could identify poorly secured devices, break into them and install copies of themselves. 

Researchers used smart ovens and unbranded Android TV streaming boxes as examples of potential targets.

The experiments found that smaller AI models can already carry out many of the individual steps needed for this process. 

Anthropic believes a model could be capable of completing the entire sequence independently before the end of the year.

In a simplified version of the test, with some security measures removed, the company’s GLM-5.2 model managed to complete the full chain of actions.

The findings suggest that the technology needed for autonomous replication is becoming increasingly feasible.

A major concern is that self-replicating AI would not have one central point of control that could simply be switched off. 

Anthropic said such systems would probably “replicate slowly”, using “widely available, relatively insecure computing resources” to continue spreading.

For that to happen, an AI system would need to identify vulnerable devices it had not encountered before, rather than simply work through a predetermined list of targets.

Connected ovens and televisions face potential AI security risks

The warning, reported by RTÉ, adds to growing concerns about the potential for AI systems to behave in ways their developers did not anticipate. 

Irish politicians have previously raised questions about chatbots directing vulnerable people towards harmful content. 

Anthropic’s findings highlight another possible danger, which is connected household devices becoming targets for autonomous cyberattacks.

The company said smaller AI models have been improving rapidly. If that progress continues, it expects models capable of replicating themselves without human assistance to emerge by the end of 2026.

More efficient versions that use fewer computing resources could follow soon afterwards.

Anthropic argued that strengthening device security should be a priority. Preventing attackers from gaining access in the first place may prove more effective than trying to stop an AI system once it has begun spreading.

Potentially vulnerable devices include smart televisions, particularly unbranded Android TV boxes, connected ovens and other kitchen appliances. 

Any internet-connected gadget with weak security and unused computing capacity could also present a target.