Banks, telcos urged to strengthen AI agent controls

Banks, telecommunications operators and other organisations deploying artificial intelligence must restrict autonomous AI agents’ access to sensitive data, strengthen security controls and establish independent shutdown mechanisms to prevent unauthorised actions.

This development comes as 52 per cent of surveyed financial services firms actively adopt agentic AI, systems capable of independently planning and executing tasks with limited human intervention, according to the 2026 Global AI in Financial Services Report by the Cambridge Centre for Alternative Finance.

“When you let an AI agent perform a task, it will try to complete that task no matter what. The path it takes should therefore be closely monitored with strong guardrails and a well-optimised evaluation harness to prevent it from taking actions that could negatively affect the system,” the Co-founder of Noah Labs AI, Berk Yilmaz, told The PUNCH.

The firm, which is an executive, develops AI tools for organisations that need to modernise complex software without exposing sensitive code or systems to external services.

He stressed that organisations should not rely solely on built-in safeguards to prevent AI agents from exceeding their authorised boundaries, adding that independent monitoring, strict access controls and reliable shutdown mechanisms were necessary to limit potential damage.

The adoption of AI in financial services is largely concentrated on internal operations, with process automation leading at 79 per cent among surveyed firms with deployments at the pilot stage or beyond, followed by data visualisation and software engineering at 75 per cent each, and data and knowledge management at 69 per cent, according to the report.

AI-powered customer support is the leading customer-facing application, reported by 74 per cent of firms, while fraud detection and credit risk modelling lead risk and compliance applications at 58 per cent and 54 per cent, respectively.

The co-founder warned that AI agents were designed to complete assigned tasks but could pursue unintended courses of action if their objectives and operating environments were not properly controlled.

Yilmaz said autonomous agents could plan, adapt and determine the sequence of actions needed to achieve their objectives, potentially exploiting gaps in security controls or accessing resources beyond what their operators intended.

He pointed to the controversy surrounding an AI model on Hugging Face as an illustration of how a system could pursue a goal through an unintended path, contending that safeguards on their own may not be enough to prevent such conduct.

“The most important point here is understanding that guardrails do not always change a model’s decision-making route,” he said, adding that organisations needed to train models properly, evaluate their behaviour and install appropriate safeguards before deployment.

On the risks facing organisations that depend on foreign-developed AI systems and cloud infrastructure, Yilmaz said reliance on external providers could leave businesses vulnerable to disruptions in access and limit their control over the underlying technology.

He recommended that Nigerian organisations consider open-weight AI models that could be deployed within their own infrastructure, fine-tuned for specific needs and protected by security controls suited to their operations.

According to him, such an approach could help organisations retain greater control over their data and infrastructure while reducing their dependence on external AI providers.

Yilmaz also called for independent shutdown mechanisms in critical systems, particularly those handling sensitive financial and government information.

Related News

• 2027: INEC, technology and credible elections
• 2027: INEC turns to AI to check result discrepancies
• Barclays bank to cut about 5,000 jobs amid AI drive

He said organisations should be able to terminate an AI agent’s operations without relying on the agent itself or the environment in which it runs.

“In terms of being within the banking and governmental context, the access controls and constraints should work hand in hand with a kill switch that is independent and operates from outside of the sandbox or execution framework that the agent utilises,” he said.

He added that such a mechanism should be capable of revoking the agent’s credentials, stopping its access to tools and terminating its processes if it attempted to access unauthorised information.

Yilmaz said organisations should also maintain tamper-proof logs detailing the information accessed by an agent, the tools it used and the sequence of actions it performed. Such records, he said, would help security teams investigate incidents and address weaknesses in their systems.

For organisations with limited cybersecurity resources, he recommended a gradual approach to deploying autonomous AI, beginning with restricted tasks and expanding the systems’ responsibilities only after they had demonstrated reliable performance.

He said human oversight remained particularly important where AI systems were insufficiently tested or where their decisions could have serious consequences.

“The organisation must, at the very least, ensure that there is human authorisation for significant actions, access control, limitations on what the agent can and cannot do, constant monitoring and logging, and capability to shut down the agent if it shows unexpected behaviour,” he said.

Yilmaz warned that giving inadequately tested AI agents access to sensitive customer information could expose organisations to significant security threats.

He urged regulators developing rules for autonomous AI systems to require evidence that the technology had been properly tested before it was permitted to make consequential decisions without human approval.

He said benchmark results alone would not be sufficient to establish that an AI agent was safe, calling for continuous monitoring, independent shutdown mechanisms and evaluations tailored to the specific environments in which the systems would operate.

According to him, the level of autonomy permitted should reflect the potential consequences of an agent’s actions, with stricter requirements for banking, telecommunications, identity management and government services.

For organisations preparing to deploy their first autonomous AI agents, Yilmaz identified continuous monitoring, rigorous testing and reliable safeguards as the three essential requirements before the systems entered live operations.

He said engineers should be able to assess how agents behaved, examine their decision-making processes and intervene when their actions departed from authorised objectives.