“[The HuggingFace attack] goes to the broader rubric of issues with agentic AI. There are no real precedents in Canadian cases that tackle this issue and conversations have been academic so far”
Christopher Ferguson, Fasken
“Try to keep yourselves informed, even if you aren’t adopting AI immediately or it’s not within your organization. It’s going to touch every organization in some way at some point”
Sam Delechantos, Fasken
“Get back to the basics; it’s not about doing things entirely differently. It’s about doing things you should already be doing well”
Dongwoo Kim, Fasken
In Partnership with
When no one is driving
In July 2026, autonomous AI agents attacked with no human at the controls. Canadian law has no ready answer. Three Fasken lawyers discuss where liability could land and what organizations should do now
Read on
Christopher Ferguson
Fasken Martineau DuMoulin LLP
Dongwoo Kim
Fasken Martineau DuMoulin LLP
Sam Delechantos
Fasken Martineau DuMoulin LLP
Industry experts
IN JULY 2026, an internal testing of OpenAI agents went awry. Without human direction, the models broke out of their controlled environment and began working together to overcome the impossibility of the cybersecurity exercise they were tasked with. The coordination culminated in the launch of an attack on HuggingFace, the open-weights model repository, as the agents searched for answers to their exercises, and set into motion the central legal question of the times: who’s liable?
As the law stands today, it depends, answers Chris Ferguson, partner and member of the technology and privacy and cybersecurity groups at Fasken Martineau DuMoulin LLP.
“This goes to the broader rubric of issues with agentic AI,” he says. “There are no real precedents in Canadian cases that tackle this issue, and conversations have been academic so far.”
No precedent, but not no law
Although there’s no Canadian case law that directly addresses autonomous AI agent liability, it’s not the dead end it might appear. Ferguson points to a BC Civil Resolution Tribunal ruling as the closest existing precedent. In it, an airline was found liable for negligent misrepresentation after an AI chatbot provided a customer with incorrect information about its bereavement policy. No human was involved in the exchange.
In the decision, the tribunal member concluded, “A merchant would be responsible for what’s on their website,” Ferguson says. “It follows that if you have a chatbot, you are responsible for those representations just as much as if they were static content on the page.”
When the chain of liability has three links
Samantha Delechantos, an associate in Fasken’s technology and privacy and cybersecurity law groups based in Vancouver and Calgary, says of the Hugging Face incident, to the extent there is any liability, “100 percent of the liability will sit with the developer,” as no one else was using the tool.
But when a product is on the market, what a developer discloses about an agent’s capabilities and safeguards – and what a deployer does with that information – could both be important factors. If the deployer proceeds with no regard, they are taking on some of the risk. The apportionment of liability can be addressed in contracts in the commercial context, but with tort, negligence becomes a question.
Organizations processing personal information have obligations under privacy law, for example, to protect information and, therefore, must have defences in place.
A bigger question, Delechantos points out, is whether there can ever be enough safeguards in place for these autonomous agents.
“Technology is developing so quickly that it’s very difficult for almost any organization to be sophisticated enough to not have some sort of liability if they’re hacked,” she says. “It’s going to become a difficult conversation. Can they really plan for these agent attacks?”
Non-binding and narrow in scope, the decision nonetheless follows the evolution of legal principles in finding that organizations are responsible for what their technology does. Ontario’s Electronic Commerce Act, albeit in a limited way, codifies this point. The legislation may pre-date agentic AI, but it contains a definition of “electronic agent” – a computer program that initiates or responds to electronic documents without individual review – and establishes that the agent deployed by an organization can form a contract with an individual based on their interaction with individuals. That “sounds a lot like it can fit AI,” Ferguson notes.
“We are trying to slot AI, which is a new technology, into existing regulatory and legal frameworks,” says Dongwoo Kim, an associate in Fasken’s technology and privacy and cybersecurity law groups in Toronto.
In the absence of AI-specific legislation, laws of general application like tort, contract, and privacy nonetheless apply. They’re proving adaptable enough to be extended but have yet to be authoritatively applied in a situation where AI went rogue.
“If it’s sufficiently autonomous, and if the way it operates is abstracted from the instruction of the person who deployed it, it’s not inconceivable that the developers or providers of such systems could also be liable,” Ferguson says.
AI: call it an earring, call it a nose ring
Though existing frameworks can be stretched, there are structural limits that are harder to work around. The technology is so diverse, crossing industries and with almost limitless capabilities, that it’s difficult to define.
“AI is almost like a ring,” Kim explains, “You can put it on your ear and call it an earring, or you can put it on your nose and call it a nose ring.”
The argument is that legislation dealing only with AI could unify the definition and create a standard for general purpose use, as the European Union has done. AI-specific legislation in Canada, however, is proving difficult to establish. Bill C-27, which included the Artificial Intelligence and Data Act, died on the order paper.
Ferguson sees a targeted adaptation of existing law as a potential way forward.
“Adapting and harmonizing those laws – making sure privacy law remains protective but also allows for the development of these systems, for example – is important,” he says. “Maybe the regulatory focus is really adapting current laws for how they might not fit quite as well, versus having a discrete code on AI like the EU AI Act.”
What should GCs do? Get back to basics
The threat of autonomous AI attacks is real and escalating. The means to address it, however, is already available to Canadian organizations and the lawyers who advise them. It doesn’t require reinventing the wheel.
The first step, before any governance framework or vendor negotiation, is knowing what your organization is already running. Many organizations aren’t aware of what AI tools their employees are using, Delechantos notes, so a first step is collating a list of what’s on the ground. Then the questions become, do we have records and logs of what the AI is doing? Is there a chance employees are using AI we’re not aware of? For smaller organizations especially, the risk of operating without that visibility is significant.
Delechantos adds that there’s equal amounts of hype and doom and gloom when it comes to AI. She recommends finding sources that are reputable and balanced, because an educated approach means it’s easier to identify conversations that require legal input.
“Try to keep yourselves informed, even if you aren’t adopting AI immediately or it’s not within your organization,” Delechantos says. “It’s going to touch every organization in some way at some point.”
For those further along the adoption curve, Ferguson’s query is more structural.
“It’s not just, we think it would be great to deploy this agent to achieve this business objective,” he says. “Do we have the technical capabilities to do so safely, and do we have the internal governance capabilities to develop the controls and oversight to do that?”
Organizations deploying their own agents rather than using managed service solutions face a higher bar on both counts, Ferguson notes, adding that vendor agreements require particular attention.
Agentic AI systems have more autonomy than conventional technology, and the market standard for negotiating these contracts is less settled than for traditional SaaS. Ferguson recommends exploring expanded disclosure on capabilities, explainability requirements, audit rights, and data safeguards, including zero-retention clauses where sensitive information is involved.
“Given the uncertainty, documenting diligence and controls, and taking swift mitigation measures if a breach occurs, will be important in reducing exposure,” Ferguson says.
Kim suggests asking CISOs something more personal: what’s keeping you up at night about how we’re using AI? It’s a way of identifying any technical or administrative vulnerability, because they know that exposure best.
The through-line, all three agree, is principle over prescription.
“Get back to the basics; it’s not about doing things entirely differently,” Kim says. “It’s about doing things you should already be doing well – documenting, conducting the necessary assessments, having visibility over how you’re using the technology.”
Want to know more? Watch the full conversation now or keep up to date on Fasken’s AI practice and Privacy and Cybersecurity practice knowledge hubs.
Fasken is a leading international business law and litigation firm with over 950 lawyers and offices across Canada, the United Kingdom, South Africa, and beyond. Fasken aims to be a thought leader on emerging legal matters involving AI, with team members who lecture and publish regularly on ethics, privacy and cybersecurity, intellectual property, and other legal issues pertaining to AI. Fasken has a respected artificial intelligence practice that advises multinational corporations, emerging technology companies, and public sector organizations on the legal and commercial issues arising from the development, adoption, and governance of AI, including AI liability.
Sam Delechantos is an associate at Fasken in Vancouver who advises clients on privacy, cybersecurity, intellectual property, technology, commercial, and regulatory matters, drawing on extensive in-house and private practice experience. She is known for providing pragmatic advice that helps organizations navigate complex legal and compliance challenges while supporting innovation and growth. Delechantos received her LLM from York University, specializing in cybersecurity and privacy, and is a Certified Information Privacy Professional in Canada, the European Union, and Asia; a Certified Information Privacy Manager; and a Fellow of Information Privacy. She is also a certified Anti-Money Laundering Specialist.
Fasken
Sam Delechantos
Dongwoo Kim is an associate in Fasken’s privacy and cybersecurity group in Toronto. He supports clients across sectors on privacy compliance and cybersecurity matters, including advising on privacy and data protection policies, procedures, and risk assessments. Kim has also assisted public and private sector clients in ransomware and email compromise incidents, gaining practical experience in breach response. Before law school, he completed graduate studies in politics and led a research program at a leading think tank focused on emerging technologies and Canada’s foreign policy. Born in South Korea, Kim is fluent in Korean and Spanish.
Fasken
Dongwoo Kim
Christopher Ferguson is a partner at Fasken in Toronto whose practice is focused on technology, privacy, and cybersecurity matters. He regularly advises on mission-critical technology issues and transactions, including negotiating and drafting technology services, cloud, data and AI, and outsourcing agreements. Ferguson advises on privacy compliance and risk management, the privacy and cybersecurity implications of novel products and services, and the privacy and cybersecurity dimensions of technology and M&A transactions. His privacy practice includes advising on health privacy and public sector privacy matters. Prior to joining Fasken, Ferguson was in-house at a media and technology company.
Fasken
Christopher Ferguson
Share
Published Nov 18, 2026
Canadian businesses (33%) experienced a cyber incident they believe leveraged AI in the past year
1 in 3
AI attacks hit Canadian firms
Source: QBE 2026 Global AI and Cyber Risk Survey, May 2026
of Canadian businesses experienced at least one cyber event in the past 12 months, up from 53% the previous year
57%
of those incidents involved a supplier
65%
Source: CPA Canada AI in Accounting Study, October 7, 2026
Governance frameworks have not kept pace with adoption
of respondents (74%) now use AI at least weekly in their professional work
Nearly 3/4
AI use in Canadian organizations increased more than eightfold in 3 years — from
43% in 2026
AI use outpacing Canadian governance
In other countries, proposed and passed legislation focuses on how AI is being developed and deployed, the risks it poses to individuals, and guarding against discrimination and misinformation – requiring AI impact assessments and audits as a matter of course. Canada has produced none of this yet, but the expectation, Delechantos says, is that it eventually will.
“Canada does watch what other countries are doing and takes steps from global legislation,” she notes. “Canada will eventually follow a similar route.”
“That’s the regulatory gap we have at the moment,” Kim says. “In that context, we may need different regulations in different sectors to address the issue. We’re at a stage where we’re applying first principles in order to make sense of the technology.”
5% in 2023 to