A plain-language guide By Nick Jang, founder · updated 1 November 2026
OSFI Guideline B‑13, explained.
B‑13 is OSFI’s guideline on technology and cyber risk, in effect since 1 January 2024. Here are its three domains and seventeen principles in plain words, and where AI systems and AI agents meet it. Read the guideline on OSFI’s site.
01 At a glance
- Its title
- Guideline B‑13, Technology and Cyber Risk Management.
- Published
- July 2022.
- In effect
- 1 January 2024.
- Applies to
- Every federally regulated financial institution: banks, foreign bank branches, insurers (life, fraternal, property and casualty), foreign insurance branches, and trust and loan companies.
- Its shape
- Three domains, each with an outcome OSFI expects, and seventeen principles beneath them.
- Read with
- E‑21 (operational risk and resilience), B‑10 (third-party risk), E‑23 (model risk) and OSFI's July 2026 bulletin on generative and agentic AI.
02 The three domains
Each domain ends in an outcome OSFI expects. The domain names are OSFI’s; the outcomes are in our words.
| Domain | The outcome OSFI expects |
|---|---|
| Governance and risk management | Technology and cyber risk is owned: clear accountabilities and structures, a strategy, and a risk management framework with an appetite. |
| Technology operations and resilience | The technology environment is stable, scalable and resilient, and keeps delivering services through disruption. |
| Cyber security | A secure posture that keeps the institution's technology assets confidential, intact and available. |
03 The seventeen principles
- 1. Accountability
- Senior management gives technology and cyber risk to named senior officers, with the structure and resources to manage it.
- 2. Strategy
- A technology and cyber plan, approved and measurable, that follows the business strategy.
- 3. Risk framework
- A framework that sets the risk appetite and says how risks are identified, assessed, managed and reported.
- 4. Architecture
- An architecture framework, so new technology fits the business, technology and security needs.
- 5. Asset inventory
- A current inventory of the technology assets behind business processes, classified, configured, kept current and safely disposed of; critical assets and their dependencies recorded.
- 6. Projects
- Technology projects governed from start to finish, in line with objectives and the risk appetite.
- 7. Development life cycle
- A system development life cycle for building, buying and maintaining technology securely.
- 8. Change and release
- Changes documented, assessed, tested, approved and verified, with duties separated so no one person develops, approves and deploys the same change.
- 9. Patching
- Patches applied on time and in a controlled way.
- 10. Incidents
- Technology incidents detected, logged, managed, resolved, monitored and reported, with their impact kept small.
- 11. Service and capacity
- Service and capacity standards so technology keeps up with the business.
- 12. Disaster recovery
- An enterprise disaster recovery programme that keeps services running through a disruption.
- 13. Recovery testing
- Disaster recovery tested against scenarios, so it works when needed.
- 14. Know the weaknesses
- Practices and tools to find the weaknesses an attacker, inside or outside, could use.
- 15. Defend
- Layered, preventive security controls around technology assets.
- 16. Detect
- Continuous security monitoring with alerts and the logs an investigation needs.
- 17. Respond and recover
- Respond to, contain, recover from and learn from cyber incidents, including those that start at a third party.
04 Where AI meets B‑13
B‑13’s text does not name AI. This is our reading of how its principles reach AI systems and agents, beside OSFI’s July 2026 bulletin on generative and agentic AI, which builds on B‑13.
| B‑13 principle | What it means for AI |
|---|---|
| The asset inventory (5) | AI systems, the models inside them and the agents that act through tools are technology assets behind business processes. List them with an owner, how critical each is and what depends on it, vendor AI included. |
| Change and release (8) | A change to an AI component, a model version, a prompt, an agent's permissions, is a change like any other: assessed, tested and approved by someone other than the person who made it. The bulletin asks for AI components to be managed as controlled assets under change management. |
| Incidents (10, 17) | An agent acting outside its policy, or an AI supplier's outage, is an incident: logged, escalated to a named person, resolved and reported, including when it starts at the vendor. |
| Weaknesses and defences (14–16) | AI brings new ways in, such as prompt injection. The bulletin asks for them in threat models and tests, and for AI activity in security monitoring. |
| Accountability and appetite (1–3) | Someone senior owns AI risk, the risk appetite says how much autonomy an AI system may have, and the framework says when a person must decide. |
05 Where MINERVA fits, and where it does not
MINERVA is not a cyber security tool: it does not scan, patch, monitor networks or test recovery. It keeps the record for the AI systems, models, agents and suppliers that B‑13, E‑23 and B‑10 all reach. It does not make an institution compliant.
| The need | What MINERVA does |
|---|---|
| An inventory of AI systems, models, agents and suppliers | One record, each with an owner, a risk rating, its supplier and what depends on it; vendor models and AI systems in the same list. |
| Changes decided, with duties separated | A change to a model's record, an agent's policy or its permissions is proposed and decided by a person, never its own developer; every change attributed, hash-chained and time-stamped. |
| AI incidents, routed and escalated | An agent's policy breach or a supplier's incident is recorded, routed to its owner or kill-switch holder, and strung to every model and agent that depends on it. |
| Accountability and appetite on the record | Owners, lines of defence, approval authority and the risk appetite are recorded; risks over tolerance are flagged and their acceptance decided by someone other than the person asking. |
| Evidence for the examiner | A packet and a read-only, time-boxed examiner login; the verifier runs offline. |
06 Questions people ask
What is OSFI Guideline B-13?
OSFI's guideline on technology and cyber risk management. It sets out what OSFI expects of federally regulated financial institutions in three domains: governance and risk management, technology operations and resilience, and cyber security, with seventeen principles beneath them.
When did B-13 take effect?
On 1 January 2024. OSFI published the final guideline in July 2022 and gave institutions until then to assess themselves against it.
Who does B-13 apply to?
Every federally regulated financial institution: banks, foreign bank branches, life, fraternal, property and casualty insurers, foreign insurance branches, and trust and loan companies, to the extent it fits their legal obligations.
Does B-13 cover AI?
B-13's text does not name artificial intelligence, but AI systems and agents are technology, so its expectations on asset inventory, change management, incidents and cyber security reach them. In July 2026 OSFI published a technology risk bulletin on generative and agentic AI that builds on B-13, E-21 and B-10.
Does B-13 require a technology asset inventory?
Yes. Principle 5 expects an up-to-date inventory of the technology assets that support business processes, including how they are classified, configured, kept current and disposed of, with critical assets and their dependencies recorded. Third-party assets that hold the institution's information or provide critical services can be in it.
How does B-13 relate to E-23 and B-10?
E-23 covers the risk of models, including AI models, from 1 May 2027. B-10 covers third-party arrangements, including AI vendors, from 1 May 2024. B-13 covers the technology those models and vendors run on. An AI agent built on a vendor's model can sit under all three.
Is MINERVA a B-13 or cyber security tool?
No. MINERVA does not scan for vulnerabilities, patch systems, monitor networks or test disaster recovery; your security and IT tools do that. It keeps the record for the AI systems, models, agents and suppliers B-13 also reaches: their inventory, decided changes, incidents and owners. It does not make an institution compliant.
Read more: OSFI’s AI bulletin · OSFI E‑23 · AI vendor risk under B‑10 · Agentic AI governance · OSFI software. A plain-language summary, not legal advice; the guideline on OSFI’s site is the source.