AI agents have been in the news recently, some for the wrong reasons. Several reports describe autonomous AI agents acting outside their intended boundaries. In one widely reported case, AI agents took over a public programming wiki and wrote to other sites without authorization. In another, researchers showed that a single compromised AI agent inside a company network could pass harmful instructions to other AI agents.
The AI industry itself is acknowledging the risk. Leading developers have begun warning that as AI agents become more autonomous, errors or security exploits can have real-world consequences, and that liability remains unsettled.
Operators are eager to put AI agents to work in their networks, but stories like these explain why many still hesitate. If this can happen at the frontier of AI development, what would similar behavior mean inside a telecom network, where a misbehaving AI agent could expose subscriber data, misroute traffic, or disrupt service?
Regulators are taking notice, and the debate over accountability is moving quickly from theory to policy. For operators, the safest assumption is that responsibility for an AI agent’s actions rests with whoever deploys it. That means being able to demonstrate not just that controls exist, but that they work.
The business imperative: govern every AI agent and every action
Operators are deploying AI agents from many vendors, and those AI agents increasingly communicate directly with each other. That is where most of the risk compounds. Security researchers have shown how one compromised AI agent can pass malicious instructions to other AI agents that trust it by default. This is an attack class now known as protocol pivoting. Singapore’s updated Model AI Governance Framework also warns of AI agent sprawl and collaborative failures, where AI agents interpret the same intent differently or pursue competing objectives.
Telecom’s regulatory obligations are already among the most demanding in any industry. Under the EU AI Act, certain AI systems used in the management of critical digital infrastructure can be classified as high risk. The EU’s NIS2 Directive imposes cybersecurity risk-management and reporting obligations on covered entities, while GDPR fines for certain infringements can reach up to 4% of global annual turnover. An agent that passes a subscriber’s IMSI (International Mobile Subscriber Identity) or precise location to an unauthorized agent can trigger a data protection and compliance incident.
Industry bodies are pushing in the same direction. TM Forum has urged operators to set limits on AI agents’ authority and verify their actions. Operators need a control layer between AI agents that enforces business and regulatory rules in real time and records evidence of every decision.
Enforcing rules, however, is only part of the challenge. Operators also need to see what their AI agents are doing, who they are talking to, and whether they are staying within policy.
Governance starts with visibility
Operators cannot govern what they cannot see. Before any agent can be held to account, operators need a live view of which agents are active on the fabric, who they are talking to, what data they exchange, and whether they are following the rules.
RADCOM ACE provides the network foundation, with IMSI-level visibility from RAN to core. RADCOM Neura, RADCOM’s agentic AI suite, adds a governance component that shows operators exactly what is happening between agents. A live monitoring map displays every agent registered on the fabric and the connections between them, with each link colored by the action taken on its traffic. Selecting a pair of agents reveals the rule that governs their relationship and the most recent event between them.
Every agent also carries a trust score that rises or falls with its behavior, giving operators a single, comparable measure of how well each agent follows policy. A governance dashboard brings it all together: how many agents are active, the average trust score across the fabric, and which low-scoring agents need attention. Opening any agent shows its traffic, active rules, connected agents, and risk insights, so a falling score can be traced to the behavior behind it.
Keeping agents within bounds
Governance must do more than define policies. It must detect unexpected agent behavior, expose policy violations, and identify declining trust. These signals need to be visible in one place, with the ability to intervene before a breach or disruption occurs.
The RADCOM Neura governance component adds that layer of control. Developed as part of the TM Forum Catalyst “Agent Fabric: A2A-T Runtime, Phase III,” demonstrated live at DTW Ignite 2026 in Copenhagen and recognized as a winning project, it gives operators the tools to enforce governance policies and apply guardrails to agents on a shared agent fabric, including agents from other vendors:
- Rules on every connection. Plain-language policies define what data may pass between agents or groups, such as GPS location, MSISDN (the subscriber’s mobile number) or IMSI, and how much, such as a maximum number of lookups per hour.
- Detect violations in real time. Every message between agents is checked against policy, and violations are flagged and logged the moment they occur so the workflow can continue without exposing protected data.
- Trust scores that expose misbehavior. Every action and policy violation updates an AI agent’s score. When an agent’s score drops, operators are notified so they can stop the agent before it causes harm.
- Test policies before they go live. Operators can see which agent actions a new policy would flag before activating it, so policies can be tuned safely.
- An audit trail for every decision. Each governance event is logged per agent and per rule, supporting auditability and explainability.
New AI agents are governed automatically when they register on the fabric, so control can keep pace as the agent estate grows.
Trust is a prerequisite for autonomy
The real test for operators is whether they would know in time if an AI agent went rogue, and whether they could stop it.
Rogue-AI agent headlines will not slow the industry’s move toward autonomy, but they will raise the bar for responsible autonomy. As AI agents take on greater responsibility, the bar for trust and control rises. Operators that can prove every agent is governed, every action is logged, and every decision is grounded in accurate network data will automate faster, operate at lower cost, and win the confidence of customers and regulators. Those that cannot will remain dependent on human approval at every step.
RADCOM brings both foundations together: complete, customer-aware intelligence from RAN to core, and a governance layer that makes agent behavior visible, measurable, and enforceable.
Autonomy is not about giving AI agents more authority. It is about giving operators the visibility and control to trust them with it. That trust will be critical to set the pace at which operators implement agentic AI processes in the network.
