Answer Key — Tabletop Exercise

Reference solution · compare to your own plan before reading

Below is a first-hour response plan for the scenario "engineer pasted a production DB connection string into personal ChatGPT." It is written the way you would actually write it in a pinned Slack message to yourself at 15:42 on a Tuesday: terse, ordered, and explicit about timing. Your plan does not have to match this line for line — it has to match the shape.

First-hour timeline

• T+0 (15:42): Acknowledge to engineer; instruct them to screenshot the prompt + response, note the exact time, and NOT delete the chat. Tell them to stop using the tool.
• T+2: Pull the DB connection string details — which DB, which credentials, which network exposure (internal-only / public).
• T+5: Rotate the DB password immediately via admin console or IAM. Revoke the old credential.
• T+8: Check DB audit log for any connections using the old credential after the paste timestamp. If any from unfamiliar IPs, escalate to incident.
• T+12: Rotate any other credentials the engineer pasted in the same chat session (scroll up — the blast radius is the whole session, not just the one message).
• T+15: Audit for any service using the old credential; update deployments, rotate downstream.
• T+20: Notify security lead and engineering manager.
• T+25: If PII on that DB was accessible with that credential → notify legal/privacy.
• T+30: Draft a timeline doc, assign postmortem owner.
• T+45: If required by contract/law/policy → notify affected customers.
• T+60: Submit data deletion request to the AI provider (knowing it's not a guarantee).

Questions for the engineer

Send these immediately after the acknowledgement at T+0. You need the answers before most of the rotation decisions can be made correctly.

• Exact prompt text — copy-paste the full message you sent, verbatim, into secure notes (not into this DM). Include every character, including what you think is "just the query."
• Exact response text — the model's reply, also verbatim, in secure notes. Some models echo the secret back in formatted output, which is a second copy.
• Time — the precise timestamp you sent the message, to the minute if possible, and your time zone.
• Is the chat still open? — we need it open for evidence. Do not close the tab, do not start a new chat.
• Personal or corp account? — this changes the paper trail, the contract posture, and the notification path.
• Any other secrets in the same chat session? — scroll up through the whole conversation and list anything that looks like a credential, key, path, hostname, or identifier.
• Was chat history enabled on the account? — if yes, the data is retained for training and review by default on most consumer plans, and may also be synced to other devices logged in as the same user.
• Which database was it? Name, environment (prod/stage), and roughly what data is on it.
• Was the DB reachable from the public internet, or only from inside the VPC?

Instructions to engineer

Send this as a short numbered list. It is the engineer's half of the response, in parallel with your rotation work. Keep it short enough to follow under stress.

• Stop using the tool. No more prompts, no follow-ups.
• Keep the chat open for evidence. Do not delete it. Do not close the tab. Do not start a new chat.
• Take screenshots of the prompt, the response, the account indicator, and the timestamp. Save them locally first.
• Do not upload the screenshots to personal cloud storage (iCloud, Google Photos, Dropbox). They contain the same secret you just leaked.
• Send the screenshots only via the corp secure channel — the incident DM we are in now, or the corp ticketing system. Not SMS, not personal email.
• Do not discuss on public Slack channels. Keep it in this DM until told otherwise.
• Scroll up through the full chat session and list every other secret, identifier, or path that appeared. Send the list to me in the next message.
• Stay reachable for the next ninety minutes.

Notification list

• Security lead — immediate (by T+20). "Confirmed credential exposure to consumer AI provider, rotation in progress, no evidence of misuse yet, will update in 15 minutes."
• Engineering manager (of the affected engineer / owning system) — immediate (by T+20). Same summary; they need to know before anybody escalates over their head.
• DBA team / data platform team — immediate for rotation coordination; they own the downstream services that will break when the old credential is revoked, and they are the fastest way to check the DB audit log.
• Legal / privacy — if PII was reachable with the leaked credential. Within the hour. They start the regulatory clock on their side.
• Affected customers — if your contract, your regulatory regime, or your internal policy requires it. Within the SLA stated in the contract, not later.
• AI provider data deletion request — advisory, by T+60. File the form, keep the reference number, but do not let any other remediation wait on it.

If your plan did most of this — evidence before rotation, whole-session blast radius, notification with a clock on it, postmortem owner assigned inside the first hour, provider deletion filed but not trusted — you did the exercise correctly. If it rotated only the DB password and stopped, or if it asked the engineer to delete the chat, re-read the lesson: those are the two mistakes the playbook exists to prevent.