AI can help government do more, faster. But attackers have the same tools, and they are moving quickly. That was the main message at DCI’s AI Security in Government event at the Canadian War Museum in Ottawa on November 18, 2026.
Public sector leaders, security teams and technology experts spent the morning on three questions:
- What is at stake?
- What are the risks?
- What should we do next?
The day opened with a keynote from Mark Schaan, Associate Deputy Minister at Innovation, Science and Economic Development Canada (ISED). Experts from F5, HPE and JFrog followed. Here is what we learned about AI security in government.
What is at stake: trust
AI will only grow in government if people trust it. As Mark Schaan put it, Canadians must use AI to benefit from it, and they must trust it to use it.
Good pilots are not enough: Government already has AI tools that work. GC Translate has translated more than 400 million words for over 310,000 public servants. CANChat frees up an estimated 15,000 to 60,000 hours of staff time each week. But a small pilot with a hand-picked team is not the same as a tool used across government. To scale, AI needs real workflows, clear owners, trained users and proof that it improves results.
Canada needs choice and control: Mark described digital sovereignty as having options, not doing everything alone. That means knowing:
- where your data is stored
- whose laws apply to it
- which suppliers you depend on
- how you would keep running if a supplier or world events changed
The opportunity is big: the share of Canadian businesses using AI has tripled since 2024. Canada’s AI for All strategy aims for 60% within five years. Getting there safely depends on security being built in, not added at the end.
The risks: six problems teams face today
- Attacks are faster than our fixes. New AI models can find software weak spots faster than teams can patch them. A weakness that is made public can be turned into an attack within minutes. The old habit of updating systems twice a year no longer works.
- More code means more weak spots. AI tools now help write much more code than before. Most of it is built on open-source parts that someone else wrote. If one of those parts is attacked, the problem spreads to everyone who uses it.
- Staff are using AI you can’t see. People want to work faster, so they use AI tools on their own, often with work data. Speakers called this “shadow AI.” Banning AI does not stop it. People just switch to their phones or home computers.
- AI agents are a new kind of user. Agents are AI tools that take actions on their own. Soon they will outnumber staff in many organizations. Most teams can’t see what their agents are doing, what data they touch, or who is responsible for them.
- Your own AI can be tricked. A public chatbot can be fooled by cleverly worded requests, called “prompt injection.” Anything the AI can reach, an attacker may be able to pull out. Projects that skip security early often break when they grow from a few users to thousands.
- It is hard to prove you are in control. Rules are getting stricter, but proving compliance is still slow and manual. Teams often spend weeks gathering evidence for an audit. With AI writing code, it can be unclear who, or what, made a change and why.
Next steps: what your team can do now
Every speaker agreed on one point: no single product will keep you safe. Strong AI security comes from good habits and several layers of protection working together. Here is where to start.
- Start with the problem, not the tool. Ask what result you want to improve, and whether AI is the right way to get there. Judge success by real outcomes, not by how good the demo looks.
- Give staff safe AI tools instead of banning AI. Offer approved tools, set clear rules for use, and teach people what not to share. Then make sure you can see which AI tools are actually in use.
- Build security in from day one. Bring security experts in early, when choices are cheaper to change. Match the level of protection to the risk. A tool that summarizes a memo needs less than one that decides who gets a benefit.
- Treat AI agents like staff. Give each agent its own identity, an owner and only the access it needs. Keep agents in separate network zones so one problem can’t spread everywhere.
- Fix the most dangerous weak spots first. You can’t patch everything at once. Focus on the weaknesses attackers are already using. Use tools that block attacks while you wait to install a fix.
- Plan to recover, not just to defend. Assume an attack will get through. Keep backups that can’t be changed, test that you can restore them quickly, and keep them out of reach of your AI tools.
- Speed up how you make changes. Monthly change meetings are too slow for today’s threats. Build in ways to test changes safely and undo them quickly if something breaks.
- Get ready for the new Treasury Board rules. The Treasury Board Security Policy Implementation Notice (SPIN), in effect since October 1, 2026, calls for a software bill of materials (a list of every part in your software) for applications that support critical services. It also expects faster patching. Building proof of compliance into your daily work beats scrambling before an audit.
- Ask vendors for proof. When you buy AI or security tools, ask how they check their open-source parts, where your data will be stored, and how they handle problems. Claims about being secure or sovereign should be backed by evidence.
- Share what you learn. Tell other teams what worked and what didn’t. A pilot that failed can teach as much as one that succeeded.
The bottom line
No one can remove every AI risk. The goal is to understand the risks well, manage them on purpose and respond quickly when something goes wrong. Teams that do this will be able to use AI with confidence, and deliver better services to Canadians.
Thank you to everyone who joined us, to Mark Schaan and ISED for the keynote, and to our sponsors F5, HPE and JFrog.