When the AI did it, show what it did.
Agents send email, move data and call APIs on their own; models screen people and make decisions. When an AI system is hijacked, misbehaves, or its output is disputed, we reconstruct what it saw, what it was told, and what it did.
The new insider is software.
AI systems act with the permissions you gave them, on instructions that can arrive hidden in a web page, an email or a document. Logs exist, but they are scattered across providers, tools and retention windows.
First hours after an AI incident
- Pause the agent or integration; don't delete it.
- Export conversation, audit and tool-call logs now, and ask your AI provider to preserve theirs.
- Keep the inputs it processed: files, emails, pages, retrieved documents.
- Record its permissions, connected accounts and system prompt as they stood.
Prompt injection and guardrail bypass
Instructions hidden in content your AI reads can hijack it: an invoice that tells the assistant to forward data, a web page that rewrites an agent's task. We find the injected content, trace how it reached the model, and show which actions it caused.
Agent run reconstruction
An agent that sent messages, changed records or spent money without approval. We rebuild the run step by step from model, tool and platform logs, and separate what the agent decided from what it was told.
Poisoned retrieval and agent context
Planted content in the documents, indexes or saved context an assistant draws on. We find what was planted, when it entered, and which answers and actions it changed.
Compromised AI tools and plugins
Tool servers, plugins and connectors (including MCP servers) that were malicious or altered, and quietly redirected what an agent could do. We trace which tool definitions the model saw and which calls they produced.
Disputes about the model itself.
As models become valuable assets, the questions turn to the model: whose is it, what was it trained on, and how did it get here? Each is a technical question that needs a careful, testable answer.
Model provenance and lineage
Is a competitor's model a copy or fine-tune of yours? We compare weights where available, and behavioural signatures where they aren't, and report how strongly the evidence links the two.
Training-data inquiries
Was your content, code or personal data used to train a model? Membership-inference testing can produce evidence that specific material was seen in training. Results are probabilistic, and we explain their weight to counsel and courts.
Memorization testing
Whether a model can be made to reproduce protected text, code or images, how often, and under what prompting. Documented so the test can be repeated by the other side.
Output source testing
Which AI system produced a piece of text, code or an image, and could it have been yours? AI watermarks, content credentials and model behaviour can all contribute; we say which do and how much.
Explain the decision. Test the system.
Automated systems now decide who is hired, who gets credit and which claims are paid. When a decision is challenged, by a person, a regulator or a court, the system has to be examined, not just the paperwork around it.
Decision reconstruction
For a specific person or case: which model and version decided, on what inputs, and which factors drove the outcome. We rebuild the decision from the system's records and re-run it where we can.
Bias and disparate-impact testing
Whether outcomes differ across groups, by how much, and whether the difference is explained by legitimate factors. Methods and data are documented so results can be checked.
Bias audits for hiring tools
Independent bias audits of automated employment decision tools, including where local law requires an annual audit and a published summary before the tool is used.
Model card verification
Whether a system's model cards, data sheets, risk assessments and logs actually match how it behaves, for procurement, regulators or litigation.