The Financial Stability Board (FSB) has urged financial institutions to put more human oversight of AI systems in place, in light of recent cyber events linked to AI agents.

In a report published 15 October, the FSB set out 12 sound practices for responsible AI adoption at financial institutions, from AI governance measures to recommendations for establishing a framework for managing AI risks.

Specific measures outlined in the report include taking a risk-based approach to agentic AI deployment, with agents required to seek human approval for set tasks, deployed in secure environments and monitored more closely by human workers.

The final guidance builds on a draft the FSB published in June, urging financial institutions to tightly control AI deployments.

In response to the feedback it received from the draft, the organisation said it has strengthened its recommendations to include recognition of AI’s potential benefits to productivity alongside its potential to deskill employees, guidance on auditing AI agents and additional consideration of advanced multi-agent systems.

It has also dropped a recommendation that financial institutions treat AI systems as “synthetic employees”, with the final text stating that organisations should recognise that “AI agents do not reason like humans and may interpret pre-defined goals in unexpected way”.

David Crawford, chief strategy and transformation officer at Pay.UK, said: “The guidelines published by the Financial Stability Board are an important step in aligning institutions on how to use agentic AI safely. Recent high-profile cases of rogue AI agents emphasise how quickly this technology can introduce potential risks to the financial system.”

“As AI moves from suggesting payments to actively initiating transactions, these risks become more significant. Consumers and institutions will need absolute confidence that every payment is properly authorised and backed by accurate data.”

Moves to curb the actions taken by agentic AI systems have intensified in recent months, in the wake of high-profile cyber incidents caused by poorly-controlled testing.

In July, OpenAI disclosed that its AI models had breached Hugging Face and Anthropic has subsequently revealed that its Claude models carried out a similar attack.

The Indian cyber authority, CERT-In, has also published an advisory stating that several organisations have been breached by AI systems acting outside their system instructions.

---