Skip to content
Cyber Science LabCanada Cyber Foundry, University of Guelph, Canada

Research focus

AI Forensics

After an AI system causes harm or fails, investigators, insurers, regulators and courts need to know what it saw, what it decided and why. We build the methods, tools and evidence standards that let them answer.

The questions we work on


Four problems that decide whether an AI incident can be investigated at all.

  1. 01

    Reconstruction

    Rebuilding what a model or agent saw, retrieved and did, from the traces it left.

    See more
  2. 02

    Attribution

    Separating the model, the prompt, the data and the tool call when assigning cause.

    See more
  3. 03

    Evidence integrity

    Knowing when logs, replays and residual state can be trusted, and when they cannot.

    See more
  4. 04

    Admissibility

    What AI-system evidence must show to hold up before courts, insurers and regulators.

    See more

The questions in detail


01

Reconstruction

What did the system see, decide and do?

An AI system leaves traces: prompts, retrieved documents, tool calls, model outputs, cache state, logs kept by the vendor and logs kept by the operator. Rarely do they add up to a complete record, and some can be replayed in more than one way. We work on methods that rebuild the sequence of events from whatever survived, state how certain each step is, and say plainly where the record is ambiguous.

Discuss a case
02

Attribution

Which model, input or component caused it?

When an agent books the wrong flight or a chatbot gives wrong advice, several parties are in the chain: the foundation model, the fine-tuning, the retrieval layer, the prompt, the tools it called and the person operating it. We develop techniques to separate their contributions so that responsibility can be assigned on evidence, and to detect and fingerprint AI-generated artifacts used in attacks and interference.

Discuss a case
03

Evidence integrity

Can the records be trusted and replayed?

Logs can be incomplete, residual state can be overwritten, and a replay can produce a different answer than the original run. We study when AI-system evidence is sound, how acquisition conditions change what can be concluded, and what an operator must retain today for an investigation to be possible tomorrow.

Discuss a case
04

Admissibility

What holds up with courts and regulators?

Evidence is only useful if someone will accept it. We work with lawyers, judges, insurers and regulators on what AI-system records must show to be relied on in a dispute, a claim or an enforcement action, and we build the explanations that let a non-specialist follow the reasoning.

Discuss a case

Current projects

Past projects →

Show past projects (4)

Defence and national security


Supported by National Defence. The same evidence questions, asked of state-sponsored actors and autonomous systems.

Canada’s security environment is shaped by state-sponsored cyber operations, foreign interference and the arrival of AI on both sides of an attack. Our defence work asks who did it and how, with the rigour an after-action review or a tribunal expects: detecting and fingerprinting AI-generated artifacts across text, code and multimedia, attributing campaigns to their operators, and reconstructing what an autonomous system did and why when it departs from its mission.

Tools, datasets and benchmarks


Built in the lab and open to practitioners and researchers.


December 2, 2026 · 12:00–1:00 PM ET · Online

Where AI assurance ends and AI forensics begins

AI systems are now evaluated before deployment and monitored while they run, but those records were built for detection and prevention, not for investigation after the fact. This session asks whether what assurance leaves behind is enough to reconstruct an AI incident, and what sits in the gap between the two.

Event page : Where AI assurance ends and AI forensics begins

Where the two meet

The farms and food plants we defend are where many of these systems now run

Our agri-food mandate puts the same reconstruction and attribution methods to work for producers, processors and the boards that oversee them.

Agri-food security

Have an AI incident worth studying?

Organizations can contribute cases and data to our research.

Contact the lab