Cyber Intelligence Quarterly Briefing — September 2026

Cyber Intelligence Quarterly Briefing - September 2026

The second edition of TECHNATION's Cyber Intelligence Quarterly Brief brings you insights from our Cybersecurity Task Force (CTF). Ahead of Cyber Security Awareness Month in October, we asked our CTF members to share their perspectives on some of the most pressing questions facing Canada's cybersecurity landscape.

A note from the Chair of TECHNATION Cybersecurity Task Force

Kevin Dawson
President and CEO, ISA Cybersecurity and Chair, TECHNATION Cybersecurity Task Force

In the last brief I wrote that the fundamentals have never mattered more, and that the threat clock is running faster than traditional governance cycles were built to handle. Three months on, the task force has gone deeper into the finding that worried me most: the identities we are handing to machines, and how quickly we are handing them out.

The questions we explore this quarter all revolve around AI governance. I think you'll find some thought-provoking ideas here, but I hope they are action-provoking ideas too. AI evolution and adoption is expanding at a pace we've never seen before in any technology. We must respond to the challenge at once. Let's dig in:

QUESTION 1: GOVERNING THE NON-HUMAN IDENTITY ESTATE
As agentic AI moves from pilot to production, what does effective governance of that estate actually look like in practice? Where should a Canadian organization start if it cannot yet say with confidence which agents and individuals hold meaningful privileges?

Every answer started in the same place. Build a working inventory of the service accounts, API keys, machine tokens, and agent credentials that let software act on the organization's behalf, record what each one is for, and put a human name against it. From there the advice is familiar enough, covering narrow permissions, credentials that expire, a person signing off on anything consequential, and someone watching for behaviour that drifts. What has changed is the size of the population those rules have to cover, and how quickly it is growing. Members kept returning to the same blunt test: who owns this agent, what can it reach, and who can switch it off?

Expert Perspectives

Mike Carzim, MGB, CISA
Senior Manager, Governance | Global Information Security and Privacy, OpenText

Effective governance of the non-human identity estate begins with visibility, accountability, and least-privilege access. Organizations should treat AI agents, service accounts, API keys, machine identities, and automation credentials as governed digital identities rather than technical artifacts. In practice, governance requires maintaining an authoritative inventory of agents and credentials, assigning accountable owners, enforcing least-privilege access, implementing continuous monitoring and audit trails, and ensuring that high-risk actions remain subject to human oversight. For Canadian organizations that cannot confidently identify which agents or individuals hold meaningful privileges, the first priority should be discovery and inventory: establish a central registry of agents, map privileges and dependencies, identify orphaned or unmanaged identities, and then apply risk-based governance controls. The objective is to reach a state where every privileged human and non-human identity is known, owned, monitored, and auditable.

Key Takeaway: If you cannot say who owns an agent, what it can access, and who can override it, you are not yet ready to scale agentic AI safely.

Jatinder Singh Mann
Founder & CEO, Cetark Corp & Netonelabs Global Operations

Machine identities now outnumber human users in the average enterprise by roughly one hundred to one. Most of them - service accounts, API keys, tokens, agent credentials - were provisioned without an owner, an expiry, or a policy for retirement. That gap is where breaches now originate. Verizon's 2026 report found that organizations expanding AI-driven identities faced a 43% breach rate, against 11% for those that did not.

Serious governance of this estate rests on five disciplines applied consistently to identities that never sign in:

• Every non-human identity has a named human owner. Where no owner can be named, the credential is revoked.
• Least privilege scoped to a task, not a role.
• Short-lived credentials by default. No static secrets embedded in code.
• Human approval required before an agent takes a high-impact action.
• Continuous monitoring for behaviour drift, with revocation built in.

The Canadian Centre for Cyber Security co-authored the Five Eyes advisory on the careful adoption of agentic AI in May 2026. That advisory is our current baseline, and it lands on the same five points.

Where should a Canadian organization begin if it cannot yet name who - or what - holds meaningful privileges within its own environment? At the beginning. Ninety days is enough to discover what exists, classify credentials by blast radius, assign a human owner to each, and revoke what nobody claims. Everything else follows from that foundation.

We should be clear-eyed about the stakes. As agentic AI moves into production across our critical sectors, the sovereignty of AI depends on our sovereignty over the identities that act on our behalf.

Andrew Buckles
Executive Vice President, ISA Cybersecurity

Effective governance of a production agentic estate starts with a control objective framework that translates your obligations into enforceable requirements. Those obligations come from legislation and regulation (privacy law, sector-specific rules, emerging AI legislation), from business requirements, and increasingly from sovereign AI strategies that dictate where models run, where data lives, and which providers are acceptable. The framework should define control requirements for both implementation and ongoing operations, and it needs to cover controls that simply did not exist in traditional IT governance: model inventorying and decommissioning, model testing, agentic system engineering standards, agentic detection and response, and governance over the skills and tools agents are permitted to use. A framework on paper is not assurance, though. Organizations need continuous validation that controls are actually working, which means recurring AI risk assessments on agentic systems and AI red teaming to surface control gaps and system-level risks before an adversary or an autonomous agent finds them first, with findings feeding a remediation loop back into the framework.

For a Canadian organization that cannot yet say with confidence which agents and individuals hold meaningful privileges, the honest answer is to start there, because you cannot govern what you cannot see. Agents are identities, and they should be governed as such: build a complete inventory of human and non-human identities, map their entitlements, and identify which ones hold privileges that matter, such as access to sensitive data, production systems, or financial workflows. From that baseline, move toward least privilege architectures where agents receive only the access their function requires, implement identity and access control monitoring so privilege drift and anomalous agent behaviour are detected rather than discovered, and run regular access assessments to keep entitlements aligned with actual need. Privilege governance is the foundation the rest of the control framework sits on; an organization that establishes it first will find every subsequent agentic control easier to implement and far easier to prove to a regulator or board.

QUESTION 2: FINDING AND SECURING SHADOW AI
Bulletin 1 noted that many organizations cannot say where AI is already in use, what it is accessing, or where data is moving. How can security and AI leaders ensure AI adoption is not stalled but still governed and then secured?

Every response set prohibition aside, and for the same reason: banning the tools moves the activity somewhere the organization cannot see it, without reducing how much of it happens. Members read shadow AI as demand the business has not met through official channels, the same dynamic that gave us shadow IT a decade ago. Their sequence is to see it, sanction it, and sustain it. That means finding out where AI is already running and what it can reach, offering approved alternatives on terms someone has actually read covering training data, retention, and jurisdiction, and then continuing to look, because the picture changes weekly rather than annually. Success has one measure: whether the approved route is faster than the workaround.

Expert Perspectives

Mike Carzim, MGB, CISA
Senior Manager, Governance | Global Information Security and Privacy, OpenText

Security and AI leaders should focus on enabling safe adoption, not restricting it. Rather than trying to eliminate shadow AI, organizations should gain visibility into where AI is being used, what data it can access, and what decisions it can influence. This starts with maintaining an inventory of AI tools, agents, and use cases, assigning ownership, and establishing clear governance policies.

Many banks, healthcare providers, and other regulated organizations restrict access to unapproved public AI services while maintaining a catalogue of enterprise-approved AI solutions backed by contractual, privacy, and security agreements with AI vendors. Information should also be properly classified and labelled; for example, data tagged as Secret should be prevented from being uploaded to or processed by AI systems, even those that are enterprise approved.

Key takeaway: Don't start with prohibition, start with visibility. Maintain an AI inventory, provide approved AI services, enforce information classification, and implement risk-based guardrails that make secure AI adoption easier than unsanctioned AI use.

Ben Lyons
Senior Director of Policy & Public Affairs, Darktrace

Security and AI leaders should treat shadow AI as a sign of unmet business demand. Blanket bans can push activity onto personal accounts and devices, making it harder to see which tools are being used, what data they access, and where that data goes.

A better approach is to provide approved tools, clear guidance and a simple route for new use cases, supported by continuous visibility across SaaS, cloud, endpoints and internal development environments. Behavioural monitoring can then identify unusual data flows or AI activity that falls outside expected business use, helping organizations govern adoption without slowing it down.

Jatinder Singh Mann
Founder & CEO, Cetark Corp & Netonelabs Global Operations

Shadow AI is better understood as a signal than a problem. It is the same signal shadow IT sent us a decade ago: when official channels are slow, people build their own. The lesson from that era stands. Prohibition drives tools underground. Sanctioning them, with pathways that match their speed, brings them back into the light.

The instinct to freeze adoption is understandable. It is also the wrong lesson. Adoption is happening whether we govern it or not. The task is to make it visible, then to make governance the path of least resistance rather than the path of most friction.

That takes three things, done in order.

• See it. Network telemetry, DNS logs, browser controls, and cloud access brokers can surface which AI tools are in use, by whom, and against what data. Most organizations have never asked. When they do, the answer is uncomfortable - and clarifying.
• Sanction it. Offer enterprise routes for the tools that matter, with vetted terms on training data, retention, and jurisdiction. The measure of success is whether the sanctioned route is faster than the shadow one. If it is not, people will keep routing around it.
• Sustain it. AI adoption changes weekly, not annually. Continuous monitoring for new vendors, new endpoints, and behavioral drift keeps the picture current.

Canada's regulatory floor is already in place. The Canadian Centre for Cyber Security's May 2026 advisory on agentic AI, PIPEDA, Quebec's Law 25, and OSFI Guideline B-13 each apply. Nothing new needs to be invented - only enforced against a landscape that moves faster than any of them were written for.

If we want the benefits of AI at national scale, we cannot govern by prohibition. We govern by making the safe choice the easy one.

Andrew Buckles
Executive Vice President, ISA Cybersecurity

We see the same pattern in most organizations: people adopt AI faster than governance can put guardrails around it, and the instinct is often to treat governance and speed as a trade-off. In practice it's the opposite. Governance done well is what makes adoption sustainable, because the alternatives are slower: an estate built on a poor foundation eventually has to be torn down and rebuilt once it can't support what's stacked on top of it, and a single serious incident can freeze an AI program for months while trust is repaired. The key for security and AI leaders is to make the sanctioned path easier than the workaround. Internally, we've encouraged AI use through an approved LLM backed by training, an early-adopter team, and regular lunch-and-learns. Open communication and security awareness keep people comfortable asking questions rather than working in secret, which lets governance grow in step with adoption instead of chasing it.

For organizations still wrestling with shadow AI, start with visibility, because you cannot govern or secure what you cannot see. Shadow AI arrives two ways: deliberately, when someone reaches for a tool, and by default, when a vendor switches AI on inside software you already own. Each needs its own response. Map where AI is already in use, personal accounts and bundled SaaS features included; what each tool can reach and where data goes, especially anything being shared with a third-party model; and which use cases are driving the demand, so the approved path can actually meet them. That visibility work is not a delay to adoption, it is the bedrock that lets you scale AI quickly without having to stop, rip out, and rebuild later.

QUESTION 3: FROM AWARENESS TO ACCOUNTABILITY AT THE LEADERSHIP LEVEL
June's Cyber Intelligence Quarterly Brief called the leadership awareness gap underappreciated and warned against treating AI as just another technology cycle. What does mature board and executive oversight of AI security look like, and what specific questions should Canadian leaders be asking their security teams over the next 24 months?

Mature oversight comes down to accountability. Decisions that once required a human signature are now being made by models and agents, and members argue that this makes AI a change in how the organization runs rather than another technology cycle to delegate and revisit at budget time. The structures they describe are ones most boards already have: an executive accountable by name, an owner and a stated purpose for every significant AI system, and controls checked continuously rather than once a year. As for the questions our experts recommended for the security teams? Don't wait 24 months to ask them: the answers are too important to wait for.

Expert Perspectives

Jatinder Singh Mann
Founder & CEO, Cetark Corp & Netonelabs Global Operations

Mature oversight of AI security does not look like a monthly update. It looks like a board that can ask hard questions and expect precise answers.

Awareness is easy. Accountability requires that leaders know what to ask - and that they keep asking until foundations are actually being built. AI security is not another technology cycle. It changes who holds trust inside the organization. Machines and models now act on our behalf, sometimes with authority that used to require a human signature.

Over the next twenty-four months, five questions belong on the standing board agenda:

• Do we know every AI system in production, who owns it, and what data it touches?
• Which decisions have been delegated to a model, and who is accountable when the model is wrong?
• What is our current inventory of non-human identities, and how many carry more privilege than they need?
• Where is our data going through AI vendors - and under what training-data and jurisdictional terms?
• If a serious AI-related incident happened tomorrow, do we have the playbook, the disclosure protocol, and the regulator relationship in place?

If the security team cannot answer with specifics, we are compounding debt. If they can, we are building foundations.

Andrew Buckles
Executive Vice President, ISA Cybersecurity

In our experience, the boards that get this right treat AI as a fundamental change in how the organization will operate, rather than another technology cycle to delegate and revisit at budget time. Mature oversight sits inside the risk and audit structure the board already uses: a named accountable executive, regular reporting, and a clear tie to the organization's risk appetite. Underneath that, oversight should be anchored to a control objective framework that translates legislative, regulatory, and business obligations, including any sovereign AI commitments, into control requirements the security team can implement and the board can measure against. Every material AI system and agent should have a named owner with a defined purpose, and assurance should be continuous rather than annual: recurring AI risk assessments on agentic systems and AI red teaming that tests whether the controls actually hold, with findings tracked to remediation. That is the difference between building foundations and quietly compounding technical debt; debt compounds precisely where controls exist on paper but nobody is validating them.

The gap between the AI of two years ago and today is staggering, and as we look ahead to the next 24 months, these are the candid conversations boards should be having with their teams:

• Do we have a current inventory of every model and agent in use, what each can access, and a defined process for testing them before deployment and decommissioning them when retired?
• Which agents and individuals hold meaningful privileges, and are we monitoring those identities and enforcing least privilege, or would a compromise be discovered rather than detected?
• If any one system or agent were compromised, what could it reach, and who is accountable?
• Where is our data going? Does anything leave Canada or move to a third-party model on terms we have reviewed?
• Learning from the Fable 5 experience, are we building reliance on services outside Canada that may be restricted or withdrawn without notice?
• When were our agentic systems last risk-assessed or red teamed, and what did we remediate as a result?
• What evolving compliance and regulatory requirements are we facing, and how are we tracking against AI best practices, guidelines, and frameworks?

Closing Reflections from the Chair

Three months ago I named non-human identity as the finding that worried me most. This round the task force came back with a method, and it is plainer than the problem might lead you to expect: know what you have, know who owns it, keep permissions narrow, and check every so often that any of it is still true.

What makes this urgent is pace. Capability is arriving faster than governance cycles were built to absorb, and every month of waiting adds more agents, more credentials, and more connections that somebody will have to account for eventually. The estate only grows. Whatever this costs an organization to sort out today, it will cost a good deal more a year from now, and the organizations that wait will be doing the same work under worse conditions.

The encouraging part is how little of it depends on spending. Almost everything the task force recommends is governance rather than technology, which means deciding who owns an agent, being clear about what it is allowed to do, and withdrawing access nobody can account for. Even an organization that sees itself sitting below the "cyber poverty line" that we discussed earlier this year can still do the work. One member puts ninety days on a first pass: that feels within reach of almost anyone.

In particular, I encourage you to reflect on the questions that were laid out in the third section above. They are worth putting to your own security team today. The answers will go a long way toward telling you where you stand.