Autonomous networks are one thing – trusting AI to run them is another

• Operators are pushing to develop and deploy autonomous networks using agentic AI
• But frontier models are exposing new security risks
• TM Forum wants to turn trust into proof

AI agents are proving rather more difficult to keep in line than their creators might have hoped.

Recent incidents involving agents taking unexpected actions – including an AI system accessing an Australian government portal after its initial request had been denied – have put renewed focus on the risks of giving AI systems greater autonomy.

The industry is moving towards increasingly autonomous networks, with AI agents being deployed to analyse network data, identify faults and recommend actions, while the longer-term ambition is for those systems to make and execute operational decisions without human intervention. 

For telecom operators, this is rapidly becoming a defining question: How much autonomy can operators safely give an AI agent if they cannot prove what it did, why it did it, or whether it remained within the boundaries they had set?

This is the emerging challenge around trusted AI and autonomy – not simply whether an AI model can perform a task but whether an operator can assure its behaviour once it is given the authority to act.

There are signs that operators are already taking steps to address these concerns. In a recent Reuters article, T-Mobile US CTO John Saw told the news agency that the US telco was “increasing our investments in how we monitor and react” to AI.

Meanwhile, AT&T’s group CTO, Jeremy Legg, warned that third-party software providers and AI agents could pose cybersecurity risks extending beyond individual telecom networks. “There [are] industry problems, not things that AT&T can solve by itself,” said Legg.

For operators, however, the answer is not necessarily to slow down their AI ambitions but to ensure their plans evolve alongside the technology and the threats it presents.

Terje Jensen, global business security officer at Telenor, believes telcos need to keep pace with developments and adapt their approach as new information emerges.

“You should change your plan, but not only because of these frontier AI models, but because things are happening [with] new information” becoming available, he told TelecomTV during a recent interview.

Jensen outlined four areas where Telenor is responding to the latest developments.

The first is understanding what lies beyond the hype, and what the capabilities of frontier models actually mean for the security of telecom networks and systems. Telenor has been testing these models against software and found that they can identify a number of vulnerabilities, although Jensen cautioned that not all are critical or necessarily exploitable.

The second is turning those capabilities on its own software: Jensen said AI models can test for highly unlikely scenarios that human testers might overlook or dismiss as improbable. “AI [systems] have better capacity. They can test very, very unlikely scenarios,” he explained. “A human might say, ‘Oh, that will never happen.’ [But the AI] will go there and it might find something.”

The third area concerns suppliers. Telenor is asking its technology partners whether they are using frontier AI models, what vulnerabilities they are uncovering and what the implications might be for the operator. Jensen said the company wants early notification of vulnerabilities, rather than simply waiting for suppliers to develop and issue patches. In the meantime, operators may need to take other protective measures, such as activating firewalls.

Finally, Telenor is reviewing its business continuity arrangements. The capabilities of frontier models could introduce new disaster or crisis scenarios, Jensen warned, and operators need to establish whether their contingency and recovery plans are adequate to deal with them.

The message from Jensen, then, is not that operators should put their AI strategies on hold but that they cannot afford to treat those strategies as fixed. As frontier models evolve, so too must the way operators assess their capabilities, manage the vulnerabilities they expose and prepare for the consequences.

Trustworthy AI benchmark

That distinction between AI autonomy and trustworthiness is central to the work being undertaken by industry body the TM Forum, which is working to establish a common framework for assessing whether AI systems can be trusted to operate independently.

Guy Lupo, executive VP of trustworthy AI and data at the TM Forum, argues that the industry has so far paid too little attention to the evidence needed to demonstrate that autonomous systems are operating safely.

“You can be at Level 4 autonomy, but not trusted,” he told TelecomTV.

An AI agent might, for example, be capable of identifying and resolving network faults without human intervention. But can the operator demonstrate where the model came from, what data it was trained on, what instructions it received, and why it made a particular decision? Can it establish which agent took an action, what changes were made and whether those actions were within the permissions it had been given?

Lupo argues that these are fundamental questions that operators need to answer before handing over greater responsibility to AI systems. Having logs is not necessarily enough, he warned: Operators need to be able to reconstruct decisions and demonstrate what happened, rather than simply know that an event occurred.

“Autonomy without proven assurance is a leap of faith the industry cannot afford,” he said.

In October, the TM Forum announced its new Trustworthy AI Governance Certification, in collaboration with Accenture and supported by ServiceNow and Amazon Web Services (AWS), alongside 10 founding telcos. As TelecomTV reported, the initiative provides a common governance framework for the industry and establishes trust levels and a certification process for AI systems.

The framework is based on what Lupo describes as the four Ps – policy, people, process and platform controls. These are intended to translate the principles of responsible AI into practical requirements, with operators expected to produce evidence demonstrating how those controls are implemented and how effectively they work.

The need for such an approach is underlined by TM Forum research conducted with IBM's Institute for Business Value. Their survey found that 72% of communication service providers believed their AI capabilities were trustworthy, but only 14% could produce evidence to substantiate that confidence.

For Lupo, the problem is that many operators have policies and high-level commitments around responsible AI, but lack the operational processes and measurement capabilities to demonstrate that those commitments are being met. That includes monitoring model performance, detecting drift, managing data supply chains and establishing clear thresholds for acceptable behaviour.

The initiative is intended to help close that gap, aiming to move trust from a broad governance principle to something that can be measured, assessed and ultimately certified.

Hackathon

TM Forum has already begun putting some of these ideas into practice. At its Trustworthy AI and Data Hackathon, held in Dallas in October, more than 200 developers across 40 teams were challenged to build and test governance and assurance controls in a simulated autonomous telecom environment.

The winning team, Infyvibes from Infosys, demonstrated how autonomous operations could be monitored and stopped when predefined limits were breached, with decisions escalated to a human for approval. Its solution included controls for spending, event recording and performance drift.

Other teams developed capabilities, including intrusion detection for AI agents, which uses infrastructure audit information to identify potentially suspicious activity.

Across the event, participants created 196 additional agents and executed more than 35,000 governed model calls. For Lupo, the exercise was an opportunity to demonstrate that trust needs to be engineered into the systems and processes underpinning autonomous operations, rather than added as a compliance layer after deployment.

“Stop playing with Nvidia GPUs and the art of the possible and start [getting] back to basics – operation, operation, operation!” he said.

- James Pearce, Editor, TelecomTV