OSFI Issues Guidance Outlining Sound Practices for Generative and Agentic AI Risk Management
As generative and agentic artificial intelligence accelerate operational and cyber risks across financial institutions, compliance and risk professionals must review OSFI’s newly articulated sound practices to prevent supervisory control gaps and ensure enterprise resilience.
Overview of the Bulletin
On July 9, 2026, the Office of the Superintendent of Financial Institutions (OSFI) issued a Technology Risk Bulletin focusing on the implications of generative artificial intelligence (Gen AI) and agentic AI for technology, cyber security, and operational resilience. While Gen AI enables content creation modeled from large datasets, agentic AI systems utilize large language models to autonomously reason, plan, make decisions, and execute actions without human intervention.
OSFI notes that although these technologies enhance organizational productivity, they simultaneously amplify the speed, scale, and automation of operational and cyber risks, driving a structural shift in the risk landscape that challenges existing risk management frameworks.
The bulletin does not establish formal regulatory expectations or mandatory rules. Instead, it articulates sound practices that institutions may tailor and adopt in a manner proportionate to their size, complexity, and risk profile. While non‑binding, these practices can inform OSFI’s supervisory assessments and complement existing regulatory frameworks, specifically Guideline B‑13 (Technology and Cyber Risk Management), Guideline E‑21 (Operational Risk Management and Resilience), and Guideline B‑10 (Third‑Party Risk Management).
Governance and Decision Integrity
Because rapid AI adoption can outpace organizational governance structures, OSFI cautions that AI systems often operate with limited human oversight and heightened reliance on third‑party models, plugins, data, and APIs. Limited senior management literacy can lead to over‑reliance on vendor assessments, reducing internal scrutiny of AI behavior.
To address these challenges, institutions may consider strengthening senior management literacy, aligning AI strategy with risk appetite, and defining ownership across the entire AI lifecycle while establishing clear limits on system autonomy.
To maintain information and decision integrity, particularly against the risk of Gen AI producing inaccurate or misleading “hallucinated” outputs that agentic AI might autonomously execute, institutions may adopt robust data classification and provenance controls to ensure data lineage is maintained. OSFI emphasizes that AI‑generated outputs should be treated strictly as inputs to decision‑making rather than definitive outcomes. Human oversight should remain central for material or high‑impact decisions, supported by auditable documentation demonstrating how AI outputs informed final analysis and outcomes.
Want daily regulatory updates delivered straight to your inbox? Subscribe to the Reg Compliance Hub here.
Software Development and Agent Autonomy
In software development and change management, OSFI highlights that while Gen AI accelerates code generation, it can inadvertently produce insecure code derived from flawed or unsafe patterns in training data. This risk is compounded by agentic AI, which can autonomously deploy defective or vulnerable code directly into production environments without human intervention.
Because AI systems and agents require extensive access to data, tools, and systems, excessive permissions create severe vulnerabilities, including multi‑step attacks and data exfiltration through tool chaining, an AI pattern where an agent executes multiple tool calls in sequence.
To mitigate risks surrounding software changes and agent autonomy, institutions may consider safeguards such as:
- AI‑generated code validation — validating AI‑generated code for security vulnerabilities and secure‑coding infractions prior to deployment.
- Controlled AI components — treating AI components (models, data, tools, agents) as controlled technology assets subject to secure development and change‑management controls.
- Non‑human identities — assigning unique non‑human identities to AI agents, enforcing least‑privilege access, and using just‑in‑time access with short‑lived credentials.
- Tool governance — restricting tool usage through allow‑lists and API gateways and enforcing approval checkpoints for high‑impact actions.
- Agent activity logging — logging and reviewing agent activity and conducting periodic access recertifications to detect anomalies.
Cyber Security Operations and Third‑Party Resilience
Advances in AI expand the cyber threat landscape through automated phishing, malicious prompt generation, and AI‑developed exploits. Institutions can also deploy AI to strengthen security operations by detecting, preventing, and responding to threats faster than traditional methods.
OSFI encourages institutions to incorporate AI misuse scenarios, such as prompt injection, into adversarial testing and threat modeling. Security teams can feed AI telemetry into existing monitoring tools to improve interaction visibility and enable rapid, automated containment actions such as blocking access or isolating compromised systems.
AI dependencies also create operational resilience and concentration risks when vendor failures disrupt critical operations or cause correlated impacts across the financial sector. Because institutions frequently rely on third‑party models, data, and APIs, and because third‑party vendors increasingly utilize AI to deliver their own services, institutions may map internal and external AI dependencies to critical operations, test AI failure and outage scenarios, establish manual fallbacks for AI‑supported processes, and enforce contractual obligations requiring third parties to disclose if and how AI is used to deliver services.
Disclaimer: This article does not constitute legal advice. Readers should seek legal or tax advice from their professional advisers regarding their specific regulatory and legal obligations.
Want daily regulatory updates delivered straight to your inbox? Subscribe to the Reg Compliance Hub here.
