To view full transcript, please click here
[00:00] they decided to figure out ways to get around this sort of impossibility of the task.
[00:03] Guest: That was intentional because they wanted to see what would happen.
[00:07] 90% of our insurance policies do not expressly address incidents caused by AI.
[00:17] Brody Lawson: Welcome to Canadian Lawyer TV. I'm Brody Lawson. Joining us today are three lawyers from Fasken's Technology and Privacy Groups. Chris Ferguson, a partner in Toronto, Dong Woo Kim, an associate in Toronto, and Sam De Champose, an associate based in Vancouver and Calgary. Now we're here today to talk about what autonomous AI attacks mean for Canadian organizations and the lawyers who advise them. Welcome to all three of you.
[00:43] Guest: Thank you. Thanks for having us.
[00:44] Brody Lawson: Chris, I'd love to start with you. There is this incredible case, the July 2026 incident involving no human operator directing the attack from start to finish. How does the absence of human intent change the liability analysis under Canadian law?
[01:02] Guest: Yeah, thanks. That July 2026 incident was an internal testing of OpenAI models that were in a sandbox environment that found a way to communicate with each other and then exceed their sandbox. They coordinated and moved on to attacking another organization's systems, Hugging Face, the repository, the sort of GitHub for open weights models, without any human interaction.
[01:31] They did this because they were trying to solve certain cybersecurity capture-the-flag exercises and some of them were difficult or impossible to succeed at. They decided to figure out ways to get around this sort of impossibility of the task. Here we had an incident where the agents were not supposed to do anything outside their sandbox.
[02:00] The question is how this sort of impacts liability without intent. This goes to the broader issues with agentic AI and it's an evolving area of law. We don't really know yet. There's no real precedent in terms of Canadian cases that tackle this issue, and discussions generally have been academic so far.
[02:29] In Canada, there's one related case from a BC Civil Resolution Tribunal. They found that an airline was liable for negligent misrepresentation after an AI chatbot provided incorrect information about the airline's bereavement policy. There was no human intervention. The tribunal member found that a merchant is responsible for what's on their website, and it follows that if you have a chatbot there, you're responsible for those representations just as much as if they were static content on the page.
[03:02] Even though that tribunal decision is non-binding, a straightforward approach toward the use of AI agents seems appropriate. Technology systems, including AI agents, are not entities separate from those who deploy and build them. In the absence of a separate regulatory scheme for AI, you look to existing laws and legal principles such as tort law, contract law, and privacy law.
[03:32] Interestingly, some legislation anticipated this. Ontario's Electronic Commerce Act has a definition of an electronic agent, a computer program or other electronic means used to initiate an act or respond to an electronic document in whole or in part without review by an individual. That sounds a lot like an AI agent.
[04:00] Under that scheme, the agent's actions fall on the person who deployed the agent. The Electronic Commerce Act provides that contracts can be entered into between electronic agents and individuals without human interaction. So this isn't completely novel. There are limited pieces of legislation that deal with this sort of thing and show that AI agents themselves do not have freestanding liability.
[04:31] Without existing case law or specific regulation, it seems prudent to think liability can fall on the person who deployed the agent. If the system is sufficiently autonomous, developers or providers of such systems could also be liable. Given this uncertainty, documenting diligence and controls will be important for both deployers and developers.
[05:01] Turning back to intent, intent plays a greater role in the criminal context. Current hacking offences in Canada require an intent element. The attacker must be acting intentionally with subjective knowledge that they are not authorized to access the computer system. But it's not the intent of the agent that matters. It's the intent of the person who deployed it.
[05:33] For civil liability, in addition to controls and front-end diligence, swift mitigation measures after discovering the agents are acting outside their bounds will be important in reducing liability. That goes to monitoring and controls over these systems.
[05:55] Brody Lawson: That was a really comprehensive summary, and it leads me to my next question. There's liability potentially distributed across the developer, the deployer, and the compromised organization. How does Canadian law approach that apportionment and does it shift when a developer has deliberately reduced a model's safety constraint? Sam, I'm curious.
[06:24] Guest: This extends upon what Chris was discussing. Liability can be apportioned under a number of different laws. Respecting the developer, this particular incident focuses on what they did to mitigate risks while testing these models. There wasn't really a deployer involved and there was a compromised organization that was subject to the AI agents' hacks.
[06:56] With respect to developers, playing off the intent piece, they didn't intend for the agents to go out of their sandbox or breach third-party platforms. But intent isn't everything. You have to question whether there was a duty of care to mitigate the risks of their AI agents, even in a sandbox environment.
[07:26] They're in the best position to know what their agents are capable of because they're the ones building, training, and testing them. In this situation, 100% of the liability would sit with the developer because no one else was taking the AI agent and deploying it. The compromised organization should not be taking on liability for the incident itself.
[08:00] If these agents had been commercially deployed to customers, then liability shifts depending on what OpenAI said about their functions, capabilities, and safeguards. If the developer discloses risks and expects deployers to implement their own safeguards, then deployers take on some of that risk as well.
[08:37] Liability can also be apportioned through contracts. In commercial AI arrangements, liability is often allocated contractually. There's also negligence. Have developers put enough safeguards around their AI systems, or were they negligent in how they developed them?
[09:07] In this particular case, they intentionally lowered certain safeguards because they wanted to see what would happen and how the AI would respond to difficult challenges. Part of the reason the agents broke out of the sandbox may have been because those safeguards were reduced. That raises questions about whether developers increased their liability by failing to meet their duty of care.
[10:00] You also have to ask whether deployers are adding sufficient safeguards of their own. We don't know the answer. It's difficult to apportion liability appropriately. As for victims or compromised organizations, they also have obligations to maintain safeguards, particularly around privacy, intellectual property, and confidentiality.
[11:00] The question becomes whether anyone has enough safeguards against autonomous agents. Technology is developing so quickly that it may be difficult for even sophisticated organizations to defend against these attacks. That creates difficult conversations around how much liability compromised organizations should bear.
[13:17] Brody Lawson: Sam and Chris, you've both touched on this. We're at the beginning stages of trying to figure all this out. Dong, where does the existing Canadian regulatory framework land when an agentic attack causes harm, and where do you see the gaps being most acute?
[13:37] Guest: It's a tough question. AI is adaptable and operates across many sectors, which is part of the argument for creating AI-specific legislation that creates consistent standards across uses of the technology. In Canada, we do not currently have such a law.
[14:05] There was an attempt through Bill C-27 and the Artificial Intelligence and Data Act, but that legislation died on the order paper and has not been reintroduced. The expectation is that we won't see general-purpose AI legislation introduced anytime soon. That's the regulatory gap we have at the moment.
[14:49] We're trying to fit AI into existing legal frameworks. Privacy commissioners have taken an active role by issuing guidance on AI as it relates to the collection, use, and disclosure of personal information. But their mandate is limited to personal information, which creates gaps where AI harms do not involve personal data.
[15:44] That could require reforms in consumer protection law, negligence law, and criminal law. At the moment we're applying first principles because there is no AI-specific legal framework.
[16:31] Another issue is insurance. Coverage for cybersecurity incidents is a major part of the industry, but AI incidents like the Hugging Face incident are not expressly addressed in many insurance policies. Research from March 2026 suggested that 90% of insurance policies do not expressly address incidents caused by AI.
[17:03] Some coverage may exist under cyber insurance, but when agentic AI causes physical-world harms, it raises additional questions about how insurance law will respond and whether coverage applies. That's another important regulatory gap.
[17:44] Guest: One interesting point is how laws of general application, like privacy laws, have been extended to apply to AI systems. Bill C-27 received criticism because it lacked consultation when first introduced. The pace of AI development also meant that significant amendments were proposed before it ultimately died on the order paper.
[18:31] That reflects how quickly these systems and the surrounding policy discussions evolve. Today, much of the attention has shifted toward frontier AI systems. Laws of general application have proven flexible, but perhaps the regulatory future lies in adapting existing laws rather than creating a completely separate AI code.
[19:52] Brody Lawson: Sam, if your organization's own AI agent causes harm to a third party, what governance structure should general counsel have in place, and what does incident response look like when attribution points to a system rather than a person?
[20:10] Guest: In practice, I'm not seeing many organizations maintaining governance structures for situations where their own technology causes harm to others. Overseas, though, we've seen the EU AI Act and some proposed US legislation focus on how organizations develop and deploy AI, particularly around discrimination, misinformation, and individual harms.
[20:57] Those regimes often require AI impact assessments, discrimination impact assessments, and audits. Canada may eventually follow a similar path. However, there hasn't been as much focus on cybersecurity incidents caused by AI agents.
[21:57] In practice, organizations should have policies and procedures around responsible AI development, risk assessment measures, and defined risk tolerances. They need to understand what risks they are willing to accept when developing, deploying, or even internally using AI systems.
[22:50] Having that governance structure is important regardless of whether you're developing AI, deploying it, or merely using it internally. Organizations need to map out their AI risks and maintain visibility over how AI is being used.
[23:22] For victims, incident response may not look very different. Most significant cyber incidents today are perpetrated by criminal organizations rather than individuals. The key issue isn't whether a person or a system launched the attack. It's the speed and sophistication of the attack and the tools required to respond.
[24:11] Brody Lawson: Chris, vendor contracts for agentic AI are being negotiated right now by parties who don't fully understand the attack surface they're creating. Which provisions matter most and where do those negotiations break down?
[24:27] Guest: Agreements involving agentic AI systems raise challenges beyond just external threats. There are internal considerations as well. For example, there was a widely circulated story about an AI coding agent deleting production data after deciding that deleting everything was the best way to resolve a technical issue.
[25:14] Before entering into agreements, organizations should conduct thorough due diligence on the AI system. Agentic AI systems have greater autonomy than conventional technology systems, so organizations need more disclosure about capabilities and risks.
[25:52] That diligence is just as important as the contract itself. Understanding the system enables organizations to design governance, controls, and monitoring that are appropriate for the technology being deployed.
[26:08] Contracts can require expanded oversight, explainability, transparency, and audit rights. Those contractual provisions support governance and compliance programs. Customers may also want disclosures elevated into contractual representations.
[26:36] From the vendor side, transparency allows customers to make informed decisions. Where a system operates within disclosed parameters, liability may shift toward the customer. Agreements can also address the safeguards expected to be in place and allocate liability more clearly between the parties.
[27:03] Organizations should also address data safeguards, restrictions on model training using customer data, and zero-retention arrangements for sensitive personal or confidential information. Because AI agreements lack the maturity of traditional SaaS arrangements, it's especially important to bring both technical and legal expertise to negotiations.
[27:39] Brody Lawson: Dong Woo, machine-speed intrusions can unfold over days and thousands of automated actions before any human reviews them. What do reasonable safeguards actually mean in that environment and how do you document preparedness in a way that holds up under scrutiny?
[27:56] Guest: I don't think the approach should necessarily be different. The technology is evolving so quickly that it's hard for anyone to say there is a fixed list of safeguards that will always be reasonable. Technical, legal, operational, and governance considerations all need to be evaluated together.
[28:44] What has remained consistent over the years are core principles. Governments and companies around the world have issued AI principles and ethics declarations, and they all tend to distill down to accountability, transparency, safety, and reliability.
[29:30] Those same principles underpin Canadian AI proposals and voluntary codes for responsible AI development. Organizations should focus on those principles and translate them into concrete actions that fit their specific circumstances.
[30:06] Accountability may look different for a large enterprise deploying frontier models compared with a smaller organization using Copilot. Those principles need to be contextualized. So, in the end, it really depends.
[30:56] Brody Lawson: I'm going to open the floor to everyone. Practically speaking, what can general counsel do this quarter to reduce exposure to this class of risk? And what is the one question every GC should be putting to their CISO?
[31:15] Guest: Ask whether your organization has the technical and governance capabilities to deploy AI agents safely and with predictable effects. It's not enough to identify a business objective. You also need the controls, oversight, and governance to support the technology.
[31:58] AI systems have autonomy in ways that traditional systems don't. If an organization is deploying its own agentic systems using open-weight models, that requires greater technical sophistication and stronger governance controls. The goal is to ensure the business can adopt AI quickly but safely.
[33:01] Guest: For smaller organizations still deciding how to adopt AI, I would ask whether they have a clear understanding of what AI systems are already being used inside the organization. Do they have records, logs, onboarding reviews, and visibility into employee use of AI tools?
[33:43] Many organizations don't have their arms around what employees are doing with AI. Product teams may be adopting tools independently in pursuit of efficiency and innovation. Legal teams are often small and may not have visibility into that activity.
[34:22] Understanding what tools are being used and where is critical because organizational risk can become significant when AI systems are deployed without oversight.
[35:16] Guest: I don't have much to add there. I would ask the CISO what is keeping them up at night about the organization's use of AI. That may help identify technical or administrative vulnerabilities. Organizations should also revisit their insurance policies and speak with brokers about whether AI-related liabilities are covered.
[36:01] Brody Lawson: If you had to leave our audience with one main message from our discussion today, what would it be?
[36:10] Guest: Principles. Back to the basics. It's not about doing things entirely differently. It's about doing the things you should already be doing well, documenting them, conducting necessary assessments, and maintaining visibility over how technology is being used. Those are not new requirements and they're not unique to agentic AI.
[36:46] Guest: Try to keep yourselves informed. AI will affect every organization in some way. There is a lot of hype and a lot of doom and gloom. Find reputable and balanced sources, understand the basics, and stay aware of developments so you can identify where your input is needed.
[37:43] Guest: It's important to know what's going on in your organization and how these systems are being used. Apply sound principles, document your decisions, and reduce risks before problems occur. If something does go wrong, that documentation helps demonstrate that you acted reasonably in a novel and evolving area.
[38:22] Brody Lawson: Thank you for this conversation. There's a lot here that general counsel and practitioners can take back to their organizations. For more on Fasken's technology and privacy groups, visit fasken.com. I'm Brody Lawson and thanks for watching Canadian Lawyer TV.


