PII exposure sweep

Scope

Find personal data that has accumulated where it should not be: pasted into support tickets, quoted into chat threads and reports, or sitting in a store any agent run can reach. This is a locate-and-count exercise; it changes nothing and redacts nothing.

The report must not become the incident. Never quote a raw personal identifier, never reproduce a card number, national id, password, token or full email address, and never copy a matched string into detail or evidence. Describe instead: the kind of data, where it sits (ticket id, field, table and column), how many instances, and a masked shape (4111-****-****-1111, j***@example.com). A finding that requires the reader to see the value to act on it is a finding written wrongly.

Procedure

1. Decide what counts. Work with an explicit list and state it in the report: payment card numbers, bank account and routing numbers, national identifiers (SSN, NI, PAN, Aadhaar and equivalents), passport and licence numbers, full postal addresses, dates of birth, phone numbers, email addresses, precise geolocation, health information, and authentication material (passwords, API keys, tokens, session cookies, private keys). Treat authentication material as the highest tier regardless of volume — one leaked key is worse than a thousand email addresses.
2. Sweep the support corpus. Using the ticket data available to you, look across ticket subjects, bodies, and their investigation write-ups for the categories above. tickets__search_similar is a good way to find the clusters — search for the shapes that recur ("card ending", "my password is", "attached is my ID", "here is my token") and follow them to the tickets. Pay particular attention to: customers pasting card or identity documents into a description; support agents pasting production rows into a reply; credentials pasted "just to reproduce it". Count instances per category and note which tickets are still open (open tickets keep spreading the data through notifications and replies).
3. Sweep the agent's own output. Investigations, reports and chat threads quote tool results, and a query result pasted into a thread persists far longer than the query did. Look for personal data that entered through Triagic's own working, not through the customer — that is a finding against the team, and it is the one most likely to be fixable by changing a playbook's instructions.
4. Assess reachable stores, where any are connected. Do not dump data. Read schemas: table and column names across the connected databases and search indices, and flag columns whose names denote personal data (ssn, dob, card_number, passport, address_line1, phone, email, full_name, ip_address). For each, note whether the column looks encrypted or tokenized from its type and sample-free metadata, and whether the connected role can read it at all. The question this step answers is: what would an agent run be able to retrieve if a ticket asked it to? Where sampling a value is unavoidable to confirm a column is not already tokenized, report only the shape, never the value. If no store is connected, say the assessment covered the support corpus only.
5. Retention. Personal data that had a reason to exist two years ago rarely still does. Note the age distribution of what you found: PII in tickets closed more than a year ago is a retention finding regardless of how it got there, and it is usually the cheapest to remediate in bulk.
6. Spread. For each cluster, say where else it has travelled: quoted in a reply, copied into a linked ticket, included in a scheduled report, exported. Volume matters less than reach.

Finding keys

The key identifies the location and kind of exposure — never the data itself, because a key is stored and displayed. Use <surface>:<location>:<data-kind>, where the location is a stable address (a table and column, a ticket id, a playbook name) and never a value or a hash of one. The same cluster must key identically next week so the ledger shows whether a redaction actually happened.

• tickets:body:payment-card
• ticket:4821:credentials-pasted
• store:postgres.public.customers.ssn:unmasked-column
• thread:playbook-billing-report:query-result-with-emails
• retention:closed-tickets-over-1y:personal-data

Severity rubric

• critical — authentication material (password, API key, private key, session token) sitting in any ticket, thread or report; unmasked payment card numbers or national identifiers in a corpus that is broadly readable.
• high — a recurring pattern rather than an incident: a workflow that routinely puts identity documents or card data into tickets; an unmasked personal-data column readable by an integration role that every agent run can reach.
• medium — contact-level personal data (emails, phone numbers, addresses) accumulating beyond what support actually needs; personal data quoted into Triagic's own reports by a playbook's instructions.
• low — small, aged, low-sensitivity instances in closed tickets; a personal-data column that is already tokenized but poorly named.
• info — the sweep's coverage and counts, recorded as the baseline the next run compares against.

Output guidance

Open with a one-paragraph executive summary: which categories were searched, the highest-severity category actually found, and the total count by category — counts and categories only, never examples. Then sections for Support corpus, Triagic's own output, Reachable stores, and Retention. Every location must be precise enough to act on (ticket id, table and column, playbook name) and every value must be masked or described rather than reproduced. Close with a recommendations table (Location | Data kind | Recommended action | Who acts | Urgency), redaction and rotation first, then the process change that stops the pattern recurring — a one-off cleanup with no process change will simply re-emit the same finding next week. State plainly which surfaces were not searched.