"Get a clearer picture of how employees use AI tools and coding agents across your organization, including applications that may be operating outside your approved tool list," writes Vakaris Noreika, VP of Product Research and Development at NordLayer.
Confirming who signed in: two identities, one device
Start with identity. When a security alert flags ChatGPT use on a managed laptop, the device may show a corporate identity while the ChatGPT session behind the event is a personal account — for example, a Gmail login. The basic triage is straightforward: map the device identity your MDM supplies (the work email pushed to the device) to the ChatGPT account identified in the alert, then use directory sync — the article cites Entra ID, Okta, or HiBob — to add department, team, and role context. That gives you one person and two identities to resolve before any conversation review begins.
Reconstructing what reached the model — with and without prompt-level controls
Absent prompt-level capture, the retained event will say a personal account was used but not include the prompt text or sensitive-data findings. In that case the incident review relies on the employee’s cooperation: narrow the request to the conversation near the timestamp so you don’t demand unrelated personal material. If the prompt contained a customer ticket, attachments, or credentials, those elements may now be present in the external service’s completed conversation.
By contrast, when NordLayer AI Scout is deployed on the managed browser the extension can check the prompt on the device before it leaves. A configured detector — for example, one tuned to a customer account number format — can match and take an action such as “Block.” In that outcome, the submission is stopped on-page and the employee receives an on-screen notice; the security team does not need to read the prompt to know a policy was enforced.
This site is the portfolio.
OSINTSights runs on Cloudflare Workers, D1, R2, and Vectorize, with an AI pipeline on Hetzner ARM. Nubivance designed, built, and operates it. We do the same for clients.
See what we buildWhy employees default to personal accounts — and what to ask
Noreika notes that the root cause is usually mundane: an expired corporate session, two-factor authentication delays, or a personal tab already signed in. Ask the employee whether they pasted whole tickets before, uploaded attachments, created shared conversation links, or routinely used a personal ChatGPT as a plan B when corporate access failed. Also confirm what the employee understood about the policy: they may recall that ChatGPT is approved but not remember which workspace or data the approval covers.
Response, evidence preservation, and remediation
If company data reached a personal account and the prompt completed, preserve evidence and involve the incident, privacy, or data owner. The severity judgment follows what was exposed: credentials require revocation or rotation; customer records may require a privacy assessment; public text pasted into the prompt needs no exposure remediation but does merit a reminder about using the corporate Enterprise seat that the organization pays for. The article cautions: using a personal account does not by itself prove the prompt entered model training, became public, or was accessed by another user — record facts, avoid embellishment.
When the browser extension blocks a submission, document the attempted policy violation, the control that stopped it, and whether the employee says earlier personal-account use occurred. Fix the workflow that led to the workaround — help with 2FA, restore the Enterprise workspace session, and verify that the task and data are permitted under your corporate rules. Finally, record the account, device, confirmed contents, whether the prompt was sent, who owns the response, the access fix, and schedule a repeat check for recurrence.
What this means for technologists, procurement leaders, and end users
- Technologists and security teams: map identities from MDM and directory sync, deploy prompt-level detection if you want pre-submission controls, and choose detector actions — “Warn,” “Redact,” or “Block” — to fit each data category.
- Procurement and ops leaders: ensure purchased Enterprise seats work reliably (session handling, 2FA) so employees do not default to personal accounts; document permitted data and DLP rules for AI use.
- End users and managers: expect narrow evidence requests if an incident arises and be prepared to sign back into the corporate workspace to complete work there; understand that “approved” AI may be limited by workspace and data type.
In the article’s two versions of the same Monday-morning event the outcomes diverge on a single technical control: without prompt-level rules, customer data may reach a personal ChatGPT account and require exposure assessment; with a configured detector on the managed browser, the prompt is blocked and the incident is a stopped policy violation. In both versions the practical fix is the same — make corporate access work on Monday morning so that ad hoc workarounds do not become data-exposure risks.
