Above Security is the inaugural sponsor of the Insider Threat Matrix
Read more
Use case
The hardest personal-AI case isn’t a one-off paste. It’s the persistent custom GPT — trained on proprietary documents, owned by the employee, and walking out with them when they leave.
Lasker beat better technicians by playing the player, not the position. He knew what his opponent really wanted on the board before they did — because he understood why they wanted it. The pattern of moves told him everything.
See custom GPTs and personal AI projects trained on your documents — by name, by content, by upload count.
One project: 30+ proprietary files, ~13.2 MB, named after the employee's own startup. Surfaced before launch.
Distinguish a productivity workaround from intentional trade-secret retention — by pattern, not paranoia.
You're responsible for IP protection at a company whose people have discovered they can build a personal Claude project or a custom ChatGPT GPT in fifteen minutes, train it on the documents they touch every day, and use it during their work. Most of them treat it as a productivity hack. A few treat it as the asset they're taking to their next role.
This is personal-AI’s harder cousin. A single paste leaves a trace in chat history; a custom GPT trained on a corpus of proprietary documents is a weaponised, portable, persistent version of the company’s IP — the moment it exists, ownership has effectively transferred to the employee who built it. Existing DLP doesn’t see it. Existing AI policy didn’t anticipate it. The investigation surfaces it before it walks.
01
Above's investigative agents surface persistent custom GPTs and named Claude projects on personal accounts, tie them to the documents that were uploaded into them, and name the employees who built them — so the inventory of portable shadow AI inside your org is a list, not a guess.
02
A single document upload is ambiguous; fifty documents uploaded into a project named after the employee's own startup is not. The investigation reads the pattern — the project name, the documents uploaded, the access cadence — so the productivity case and the IP-retention case never share a default response.
03
When a confirmed custom GPT is trained on trade-secret material, the investigation hands legal the document list, the platform, the dates, and the employee — so the trade-secret conversation begins with evidence, not a memo describing what someone thinks they saw.
Most personal AI use is a productivity hack — and most of it isn't insider risk. What turns it into insider risk is persistence and portability: a custom GPT trained on dozens of proprietary documents is a weaponised version of the company's IP. The employee owns the model, the model contains the documents, and both walk out the door when the employee leaves. The insider-risk umbrella covers negligent and intentional; in this pattern, the act of building a persistent personal asset on trade-secret material crosses the line regardless of intent at the moment of upload.
DLP fires on data movement against predefined policies. Uploading a document into a personal ChatGPT project is, mechanically, the same as opening it in any tab — there's no policy event the rule was written for. The threat isn't the upload; it's what the upload creates. A persistent personal AI trained on dozens of proprietary documents is an asset that didn't exist before, and DLP categories don't reason about created assets. It requires continuous behavioral investigation that tracks the relationship between the uploaded documents, the AI project, and the employee over time.
Above's investigative agents surface the persistent custom GPT, name the documents uploaded into it, identify the platform and account type, and link it to the employee who built it — alongside the behavioral context (project name, access cadence, peer activity, departure signals). The output is a structured investigation: timeline, contextual analysis, reasoning, recommended actions. For trade-secret material, the package is ready for IP counsel. For lower-risk productivity-only projects, the recommendation is education, not escalation.
