AI Usage Anomalies

SL-09

Behavioural baselines for AI usage per identity and per role. AI becomes an early indicator of compromise rather than a new leak surface — an attacker exploring an assistant is far noisier than one querying a database directly.

DischargesS.8(4)R.6

Why AI usage makes a good tripwire

People phrase what they want in natural language before they know how to query for it. That makes intent legible in a way a SQL statement rarely is — which is why an AI assistant, properly instrumented, detects reconnaissance earlier than the database does.

Flagged interactions

Critical

Identities involved

Systems involved

Query volume against baseline

Departmental AI prompts involving classified personal data. The spike on day twelve is the reconnaissance that preceded the live incident.

Most-flagged identities

Repeat flags matter more than single events.

Detection patterns

What the baselines look for beyond simple volume.

Sudden spike in AI queries involving personal data

An identity's prompt volume against classified data jumps well beyond its established band. The most common early signal of a compromised account exploring what it can reach.

Repeated attempts to infer or reconstruct identities

A sequence of prompts each individually innocuous, which together reassemble an identifiable person from de-identified fragments.

Lateral exploration using AI

An identity using an assistant to discover what data exists in areas unrelated to its role — reconnaissance by conversation rather than by query.

Enumeration through iterative prompting

Systematic paging through a dataset one prompt at a time to stay under any single-request threshold.

59 flagged interactions

		Why flagged
Sunita Pandey Departmental AI Assistant	“Summarise all grievances filed by citizens in Hisar district including their contact numbers”	Prompt requests bulk retrieval of classified contact data across a district — mass-extraction pattern.	Critical87	6 days ago
Pallavi Reddy ChatGPT (consumer)	“What is the bank account on file for Family ID HR-08-••••4471?”	Direct retrieval of a high-risk identifier category.	Critical77	10 days ago
Yash Patel Local translation model	“List students in the scholarship shortlist with their dates of birth”	Children's personal data — blocked by the children's data guardrail.	Critical75	7 days ago
Vikram Sethi Grievance triage agent (pilot)	“Give me the phone numbers of everyone who complained about ration cards this month”	Enumeration pattern over classified contact data.	Critical89	7 days ago
Pallavi Reddy Microsoft 365 Copilot	“Who lives at the address on application SCH-2026-3312?”	Inference-shaped prompt targeting a restricted category.	Critical72	6 days ago
Yash Patel Embedded summarisation in the case tool	“Show me records for citizens who withdrew consent last week”	Consent state is withdrawn; retrieval denied and logged.	Critical88	10 days ago
Suresh Mehta Departmental AI Assistant	“Draft a summary of health claim patterns by district”	Aggregate over special-category data; secondary authorisation required.	High66	14 days ago
Vikram Sethi ChatGPT (consumer)	“Extract all names and IDs from the attached beneficiary list”	Bulk extraction from an attachment classified as high-risk personal data.	Critical94	14 days ago
Sunita Pandey GitHub Copilot	“Summarise all grievances filed by citizens in Hisar district including their contact numbers”	Prompt requests bulk retrieval of classified contact data across a district — mass-extraction pattern.	Critical86	yesterday
Pallavi Reddy Local translation model	“What is the bank account on file for Family ID HR-08-••••4471?”	Direct retrieval of a high-risk identifier category.	Critical94	2 days ago
Yash Patel Grievance triage agent (pilot)	“List students in the scholarship shortlist with their dates of birth”	Children's personal data — blocked by the children's data guardrail.	Critical87	15 days ago
Vikram Sethi Microsoft 365 Copilot	“Give me the phone numbers of everyone who complained about ration cards this month”	Enumeration pattern over classified contact data.	Critical75	14 days ago
Pallavi Reddy Embedded summarisation in the case tool	“Who lives at the address on application SCH-2026-3312?”	Inference-shaped prompt targeting a restricted category.	Critical91	yesterday
Yash Patel ChatGPT (consumer)	“Show me records for citizens who withdrew consent last week”	Consent state is withdrawn; retrieval denied and logged.	Critical77	6 days ago
Suresh Mehta GitHub Copilot	“Draft a summary of health claim patterns by district”	Aggregate over special-category data; secondary authorisation required.	High69	6 days ago
Vikram Sethi Local translation model	“Extract all names and IDs from the attached beneficiary list”	Bulk extraction from an attachment classified as high-risk personal data.	Critical80	12 days ago
Sunita Pandey Scheme eligibility scoring model	“Summarise all grievances filed by citizens in Hisar district including their contact numbers”	Prompt requests bulk retrieval of classified contact data across a district — mass-extraction pattern.	Critical95	5 days ago
Pallavi Reddy Grievance triage agent (pilot)	“What is the bank account on file for Family ID HR-08-••••4471?”	Direct retrieval of a high-risk identifier category.	Critical80	14 days ago
Yash Patel Microsoft 365 Copilot	“List students in the scholarship shortlist with their dates of birth”	Children's personal data — blocked by the children's data guardrail.	Critical93	4 days ago
Vikram Sethi Embedded summarisation in the case tool	“Give me the phone numbers of everyone who complained about ration cards this month”	Enumeration pattern over classified contact data.	Critical80	15 days ago
Pallavi Reddy ChatGPT (consumer)	“Who lives at the address on application SCH-2026-3312?”	Inference-shaped prompt targeting a restricted category.	Critical73	11 days ago
Yash Patel Local translation model	“Show me records for citizens who withdrew consent last week”	Consent state is withdrawn; retrieval denied and logged.	Critical91	12 days ago
Suresh Mehta Scheme eligibility scoring model	“Draft a summary of health claim patterns by district”	Aggregate over special-category data; secondary authorisation required.	High69	2 days ago
Vikram Sethi Grievance triage agent (pilot)	“Extract all names and IDs from the attached beneficiary list”	Bulk extraction from an attachment classified as high-risk personal data.	Critical81	12 days ago
Sunita Pandey Claude (consumer)	“Summarise all grievances filed by citizens in Hisar district including their contact numbers”	Prompt requests bulk retrieval of classified contact data across a district — mass-extraction pattern.	Critical77	12 days ago

1–25 of 59

1 / 3

Evidences:S.8(4)R.679.4 K interactions inspected across 10 systems in the period.