Featured publication · AI agent incident investigation
After the Agent “Escaped”: Investigating the Investigation
The July 2026 OpenAI and Hugging Face incident shows why an agent's actions and an operator's response must be reconstructed together.
Zemi North / Insights
Research for claims that must survive review.
Analysis of digital evidence, AI-influenced decisions, and the controls required to examine them.
Digital evidence
AI assurance
Decision reconstruction
Current record
Published analysis
Each publication establishes its scope, identifies the evidence required, and marks the limits of what the available record can support.
Publication index
Released analysis
09 recordsAfter the Agent “Escaped”: Investigating the Investigation
The July 2026 OpenAI and Hugging Face incident shows why an agent's actions and an operator's response must be reconstructed together.
Before the Examination: What Digital Evidence to Preserve
A practical guide for counsel, employers, and incident managers on preserving devices, cloud records, logs, and decision context before a forensic examination begins.
Investigating AI-Enabled Ransomware: When the Attack Chain Includes a Model
AI-mediated malware can move part of an attack chain into prompts, model responses, generated code, and volatile memory. A synthetic ransomware investigation shows how to reconstruct what the model did.
Who Did It: The User or the Agent?
An AI-mediated data disclosure shows why task initiation, technical execution, intent, authorization, control, and accountability require separate findings.
The Replit Deletion Is Old. Its Evidence Problem Is Not.
What a widely repeated AI incident establishes, what it does not, and why the distinction still matters.
Claude Left a Mark. What Can Investigators Prove?
What Claude's text watermark can establish, where provider records may extend the trail, and why attribution to a person still requires corroboration.
AI Assurance as the Connective Layer
How evaluation, validation, red teaming, governance, monitoring, and audit combine to support bounded assurance findings about deployed AI systems.
Provenance Overreach: When an Artifact Proves Less Than It Appears To
A forensic examination of why agent, account, and token records do not automatically establish human authorship, intent, authorization, or control.
What Evidence Must Be Preserved When AI Influences a Decision?
An evidentiary guide to the prompts, data, system state, logs, human interventions, decisions, and actions needed to reconstruct AI-influenced decisions.
Research foundation
The Zemi Method
Insights apply the evidentiary discipline. The Method defines it, versions it, and keeps the reasoning open to examination.
Publication programme
Questions under examination.
- 01
Preservation, integrity, and provenance in AI-influenced decisions.
- 02
Control and attribution when AI systems execute consequential actions.
- 03
Forensic limits in synthetic-media and automated-system claims.