StarHub to trial SIM-based IDs for governing AI agents

As artificial intelligence (AI) agents increasingly operate behind the scenes on smartphones and industrial machines, Singapore telco StarHub is developing a world-first SIM-based governance service to act as a real-time kill switch against rogue AI behaviour.

The ambitious project, which is currently in the proof-of-concept stage and slated for trials by mid-summer, aims to tackle the growing cyber security blind spot posed by AI agents powered by advanced AI models to make decisions and execute tasks autonomously.

Speaking in a recent interview with Computer Weekly, StarHub’s chief technology officer Volkan Sevindik warned that consumers would be largely unaware of the sheer volume of AI agents operating in their hardware.

“In the near future, there will be maybe more than 1,000 agents running on a phone to perform many different tasks so that your life becomes easier,” said Sevindik. “The risk is, what are these agents doing in real-time in the background?”

Without proper safeguards, Sevindik cautioned that AI agents could quietly access personal identifiable information (PII), such as cloud photos or banking details, to train external models or share sensitive data with unknown third parties.

To mitigate the risk, StarHub is building a trust layer that leverages the permanent, read-only nature of a mobile SIM card to assign unique, real-time IDs to every AI agent running on a user’s device. This will expose the IPs that agents are communicating with, the data they are sharing, and the prompts used.

If an AI agent begins behaving maliciously, StarHub can cut its connection instantly. The telco employs various sizes of orchestrator models to operate guardian agents that monitor the activities of AI agents across its mobile, broadband, and TV networks. “We will never understand AI with human perception. We can only understand AI with another AI,” Sevindik noted.

Beyond consumer smartphones, the enterprise applications for SIM-based governance are significant, particularly for critical infrastructure. Sevindik highlighted use cases spanning autonomous vehicles to robotic surgeries and pharmaceutical manufacturing.

“If one of the manufacturing robots, which will have agents running in them, goes down, the whole production line will go down. There could be billions of dollars of losses,” he said. “As a trusted telecom partner, we can sense any wrong behaviour in advance and cut the connection between robots, so it doesn’t impact the manufacturing line.”

Because no system is foolproof, StarHub is also in talks with insurance partners to underwrite the service. The goal is to commercialise the service as an “AI Trust Package”, which provides users and enterprises with financial protection in the event of a data leak by an AI agent, for example.

To support the massive data requirements of agentic AI, StarHub is already shoring up its network infrastructure. While traditional internet usage heavily relies on downloading data, AI agents constantly upload information to communicate with each other and perform tasks.

However, Sevindik noted that networks are traditionally designed for downlink and not uplink traffic, which AI agents will increasingly account for. “We are re-optimising our network based on uplink and changing aspects of the network to be more balanced,” he added.

StarHub is currently engaging with the Infocomm Media Development Authority (IMDA), which has shown interest in the service, Sevindik noted.

“Everyone believes [agentic AI] works and there will not be any problems, but I’m not sure if that would be the case, because the number of agents is increasing significantly," he said. “We want to provide that trust layer to protect you without you knowing it.”