Anthropic · Safety and security · Research
Anthropic says small AI models may be months away from being able to self-replicate on IoT devices
In a report published October 12, Anthropic’s Frontier Red Team wrote that “small AI models may be months away from being capable enough to self-replicate” on commercial internet-of-things devices such as smart ovens and off-brand Android TV boxes. The report studies worms that run their own model directly on compromised hardware, without a datacenter.
In a firewalled test network of eight emulated Android nodes, GLM-5.2 was the only tested open-weight model capable of full end-to-end replication, reaching all eight nodes in four of five trials. Models of 30 billion parameters and under completed many individual steps but made mistakes when chaining them. Anthropic says small models will match that level by the end of 2026 if they keep improving at their historical rate.
Anthropic notes that the evaluation kept the cyber-exploit step trivial, making the results an upper bound, and that it does not yet have a full rogue replicator.
Sources
- Anthropic, “AI models may be capable of self-replicating without access to a datacenter”, Frontier Red Team report (William Zhou, Daniel Freeman), (anthropic.com)
- SecurityWeek, “Small AI Models Could Soon Self-Replicate on IoT Devices, Anthropic Warns”, Summary of the report's findings and limitations, (securityweek.com)
About this story
Spotted an error in this story? [email protected] · Instagram
Short link: thejarrus.com/en/iot-self-replication
This story in Turkish: Anthropic, küçük yapay zeka modellerinin IoT cihazlarında kendini kopyalayabilmeye birkaç ay uzakta olabileceğini açıkladı
On the same topic
NTT Data is in the final stages of talks to build Anthropic’s first data center in Japan, Nikkei reports
Nikkei says NTT Data is close to a deal to build an Osaka data center for Anthropic with up to 500 MW of power and construction costs of up to ¥1 trillion.
Debate over OpenAI’s math results grows as Recht pans its FFT claim, Hales warns about Lean and Yau launches an AI math alliance
Ben Recht called OpenAI’s FFT result an exploit, Thomas Hales warned against blind trust in Lean, and Shing-Tung Yau launched an AI math alliance in China.
Anthropic says its new distillation defenses have cut the data harvested from Claude by around 90 to 99 percent
Anthropic says preserved thinking, restricted reasoning access and a two-stage classifier have cut the data distillers harvest from Claude by 90-99%.