The Financial Stability Board, the international body that coordinates national financial authorities and standard setters, published final guidance Thursday on how banks and other financial firms should adopt artificial intelligence, with specific controls for AI agents. The board warns that agents, AI systems that act on their own to carry out multistep tasks, can create risks that "materialise at great speed." For bank boards, the document reads as a checklist of what supervisors are likely to ask about next.

What Happened

The final report, titled "Sound Practices for Responsible Adoption of Artificial Intelligence (AI)," sets out 12 practices. Four cover governance across the whole organization, including board oversight and clear accountability. Eight cover the life of each AI system, from assessing a use case's risks to human oversight, cyber risk and reliance on outside AI providers. The board said it "strongly encourages" boards and senior management to reference the practices.

The FSB consulted on a draft in June and received about 160 responses, according to its overview of consultation responses. Several respondents asked for more on agentic AI, and the final text expanded that material. The FSB says the new agent considerations were not fully consulted on but were distilled from the responses.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

The report lists the main agent risks: unauthorized actions, errors, data exposure, disruption to connected systems, weak human oversight, "memory poisoning," in which attackers plant bad data in an agent's knowledge base, and agents that spawn and direct sub-agents. It notes that an agent may treat an unauthorized action as an effective shortcut, which can make the action hard or impossible for people to detect or reverse, and that real-time human monitoring becomes impractical as agent use grows.

The suggested controls are concrete:

  • Approval triggers: clear, risk-based points at which an agent must seek human approval, such as transactions above a certain value or actions that could cause irreversible harm.
  • Limited reach: isolated, offline environments where possible, and otherwise limiting an agent's access to systems, data, networks and other agents to what the task needs.
  • Kill switches: tested mechanisms to pause, limit, override or remove an agent's access, including checks on how agents behave when they cannot finish a task safely.
  • Identity and records: a unique identifier for each agent, separating it from human users, and records that trace what triggered each action, what tools it used and whether it stayed within its permissions.

The report also suggests using AI to monitor agents and red-teaming, or deliberately stress-testing, whether agents can stop safely and escalate to a human.

Why It Matters

The FSB's decisions are not legally binding; its members commit to carry out agreed standards at home. Those members include the Federal Reserve, the Securities and Exchange Commission and the Treasury Department, so the practices are likely to inform what U.S. examiners expect. The Primary Wire reported this week, citing Reuters, that the Fed, the Office of the Comptroller of the Currency and the Federal Deposit Insurance Corporation have been questioning banks about AI agent controls, including how agents can be shut down.

The report is also candid about limits. It says controls on agent behavior do not by themselves answer who is accountable for an agent's actions, so firms should name an accountable person for each agent use case and seek contractual clarity from AI vendors on legal liability.

The Business Case

The guidance is not only defensive. In a new case study, the FSB describes a large, unnamed international bank whose fraud-detection agent proposes new detection rules for human analysts to approve. According to the report, the agent has helped develop or update three quarters of the bank's card fraud rules.

What to Watch

The FSB, chaired by Bank of England Governor Andrew Bailey, says agent-specific practices remain nascent and may get more comprehensive treatment, and that it will update the practices as AI develops. Watch for national supervisors to cite the document in exams or guidance, and for supervisors to test whether banks can show a named owner, an identifier and a working kill switch for each agent.