IOSCO’s New Toolkit: Financial AI Supervision Shifts Towards Continuous Testing

IOSCO’s New Toolkit: Financial AI Supervision Shifts Towards Continuous Testing

A New Era in Regulation: IOSCO Provides a Practical Framework for Supervisors

The International Organization of Securities Commissions (IOSCO), the global standard-setter for the securities and capital markets sector, has unveiled a new supervisory toolkit for artificial intelligence. The organization, which brings together the world’s national securities regulators like the SEC in the United States and the FCA in the UK, is sending a clear signal to the market: the deployment of AI systems must be governed by strict frameworks for governance, testing, and lifecycle oversight. According to Jean-Paul Servais, board chair of IOSCO, “The increasing integration of artificial intelligence into capital markets requires supervisors to have practical and proportionate tools to assess emerging risks while supporting innovation and safeguarding market integrity and investor protection.”

The toolkit, the result of two years of work, covers the full AI lifecycle, from traditional machine learning models to Generative AI and emerging agentic architectures. It is explicitly designed for use during “on-site examinations and inspections.” The Madrid-based organization noted that “the use of AI in capital markets presents opportunities and risks that supervisors must understand and address to protect investors, maintain market integrity, and preserve financial stability.”

Do you have a question about AI security? You can reach us here:

From an AIQ standpoint, this move has implications far beyond the financial sector. In the European regulatory landscape shaped by the EU AI Act and GDPR, this toolkit can serve as a blueprint for the specific, auditable requirements that authorities will impose on high-risk AI systems. The toolkit’s requirements for “AI system decision and usage logs and audit trail documentation” and “validation and testing policies & procedures and reports” directly mirror the AI Act’s transparency and documentation obligations.

Specific Risks of Generative AI in the Regulatory Spotlight

The IOSCO report pays special attention to the challenges posed by Generative AI and agentic systems. The organization states that survey data indicates “there is an increase in the speed and scale of AI development and implementation over the last two years, particularly using GenAI.” In parallel, however, they acknowledge that “GenAI systems may be difficult to predict, evaluate, understand, explain, and test.”

The document highlights several specific vulnerabilities, including hallucinations, model drift, adversarial attacks, data poisoning, bias, and AI system failures. Hallucinations are defined as “outputs that appear plausible but are factually or logically incorrect or fabricated.” While the report outlines mitigation techniques such as Retrieval-Augmented Generation (RAG) or Chain-of-Verification frameworks, it cautions against overconfidence: “these techniques have limitations and are unlikely to fully eliminate the risk of hallucinations.”

In a corporate context, this means that risk management must focus on the threats identified by the OWASP LLM Top 10. The data poisoning (LLM-03: Poisoning), system failures and unexpected behaviors (LLM-06: Overreliance, LLM-10: Unsafe Output Handling), and hallucinations (LLM-09: Model Hallucination) mentioned by IOSCO are all areas for which a responsible organization must have a dedicated testing and validation strategy. The toolkit effectively builds a set of regulatory expectations around the technical risks identified by OWASP.

The Supervisory Expectation: Continuous Testing is the New Normal

Perhaps the most critical message of the IOSCO toolkit is the necessity of continuous oversight and testing. The document makes it clear that regulators expect pre-deployment, post-deployment, and ongoing validation, monitoring, and testing. The report repeatedly warns that supervisors may increasingly expect firms to demonstrate not just that AI systems work, but that they can be governed, explained, and interrupted safely when necessary.

Supervisory concerns also extend to complex risks like “unexpected emergent behaviors” or the “potential for cascading impacts or failures across interconnected systems.” The report suggests that firms may need continuous monitoring, gradual rollouts, resilience and red team testing, rollback plans, and human oversight configurations. Furthermore, “AI-enabled cyber capabilities” could “materially accelerate threat evolution and increase the speed, scope, and scale of existing attack techniques.”

The AIQ perspective is that this represents a paradigm shift. The “deploy and forget” model is obsolete. AI systems, especially LLMs, are dynamic; their behavior can change over time (model drift). This means that compliance—whether with GDPR data protection principles or EU AI Act requirements—is not a one-time state but a continuously maintained property. Red teaming, as a method for proactively discovering vulnerabilities in AI systems, moves from the “nice-to-have” category to “must-have” with this regulatory expectation, especially in critical financial and other regulated industries.

Implications for the Hungarian and EU Markets

Although the IOSCO toolkit directly targets capital market participants, its message is universal. For technology and security leaders, the key takeaway is that AI governance is no longer a theoretical issue but a mandatory, continuous discipline of testing and validation that regulatory authorities will actively monitor and enforce.

According to Hanzo van Beusekom, Chair of IOSCO’s Fintech Task Force, the toolkit “provides practical guidance to help authorities address the risk arising from the growing use of AI systems across financial markets.” This proactive regulatory stance suggests that with the EU AI Act coming into force, European companies developing or deploying high-risk systems can expect similarly detailed inspections and expectations. Preparation must begin today by building robust governance frameworks, developing continuous testing capabilities, and implementing independent AI security audits.

Attila Rácz-Akácosi

Independent AI Security Specialist

Two decades of analytical and systems-oriented experience. I have been working with artificial intelligence since 2017. In recent years, I have specialized in AI/LLM security and AI Red Teaming. Systems-level thinking instead of endless vulnerability checklists.