DLP flagged customer records pasted into an unsanctioned public AI chatbot

1. Triage and prioritization. This is a confirmed data-exposure event, not a malware intrusion, so the priority is scope determination and regulatory notification rather than eradication. Highest priority: confirm exactly what left the environment (42 rows of customer PII plus an internal script) and whether the vendor retains it. Treat the customer PII as the high-value asset because it is regulated data.

2. Evidence preservation first, before touching the endpoint. Export and hash the DLP match record, the full web-proxy log for JORDAN-WKS around 13:46-13:55 UTC, and the endpoint agent clipboard event. Preserve browser history, profile data and any local cache on JORDAN-WKS via a forensic copy before the user logs in again, since a reimage or a browser cache clear would destroy the only client-side record of the prompt text. Record chain of custody with timestamps and who collected what. Do not have Jordan delete the chat thread yet, because that destroys evidence of scope.

3. Containment. Block chat.example-ai[.]com at the proxy and enable the Generative AI category block that was missing, so further pastes cannot occur from any host. Verify no other users have posted to the same domain in the last 30 days by querying proxy logs for that category. Do not disable Jordan's account immediately: this appears to be a well-intentioned policy violation, not a malicious insider, and locking the account destroys cooperation without reducing risk. Instead, temporarily restrict Jordan's access to the customer-export report until scope is confirmed.

4. Investigation and scope. Interview Jordan in a non-punitive way to obtain the exact prompt text and whether the chat was in a personal account with training-on-data enabled. Pull the customer-export report logs to identify precisely which 42 customer records were included (names, emails, phones, account ids). Review the internal script to determine whether it contains credentials, API keys or internal hostnames; if it does, treat those secrets as compromised and rotate them.

5. Recovery. Rotate any credential or key found in the leaked script. Request deletion of the conversation and data from the AI vendor, and document their response, noting that with no DPA and no enterprise retention control we cannot rely on deletion. Notify Legal and Privacy immediately so they can assess breach-notification obligations for the 42 affected individuals under applicable regulation, and notify the data owner and the Support team manager.

6. Hardening and follow-up. Add sanctioned AI tools to the app catalog with SSO, publish clear shadow-AI guidance, add DLP blocking (not just alerting) for PII paste into GenAI categories, and run awareness training for Support. Write a post-incident report with timeline, scope, and the control gaps that allowed an uncategorized GenAI destination.