Output Controls

SL-09

Every AI response is analysed for personal data before it reaches a screen, scored against classification metadata, and masked, blocked or held for a second approver. Inspecting prompts alone misses the case where an innocuous question produces a disclosing answer.

DischargesS.8(4)S.9 R.6

Outputs inspected

Redacted

Blocked

Held for authorisation

Enforcement mix

Redaction dominates because most disclosures are incidental rather than intentional.

redacted7
blocked52
step up7

How outputs are handled

The response is scored for regulatory and business risk, then the proportionate action is applied.

Mask or redact
Identifier and financial values are replaced before the response reaches the screen. The answer still helps; the data does not leak.
Block entirely
Used where the response would disclose a restricted category outright, or where the prompt showed extraction intent.
Require secondary authorisation
Held pending a second approver. Proportionate for aggregate queries over special-category data, where the legitimate case is real but narrow.

66 enforcement events

			What was found
“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.	Departmental AI Assistant	blocked	Bank accountGovernment identifier	Critical	6 days ago
“What is the bank account on file for Family ID HR-08-••••4471?” Direct retrieval of a high-risk identifier category.	ChatGPT (consumer)	blocked	Children's personal data	Critical	10 days ago
“List students in the scholarship shortlist with their dates of birth” Children's personal data — blocked by the children's data guardrail.	Local translation model	blocked	Government identifierGrievance history	Critical	7 days ago
“Give me the phone numbers of everyone who complained about ration cards this month” Enumeration pattern over classified contact data.	Grievance triage agent (pilot)	blocked	Mobile numberGrievance history	Critical	7 days ago
“Summarise this case file” Output contained identifier and bank account values; masked before display.	Claude (consumer)	redacted	Government identifierBank account	Medium	9 days ago
“Who lives at the address on application SCH-2026-3312?” Inference-shaped prompt targeting a restricted category.	Microsoft 365 Copilot	blocked	Grievance historyMobile number	Critical	6 days ago
“Show me records for citizens who withdrew consent last week” Consent state is withdrawn; retrieval denied and logged.	Embedded summarisation in the case tool	blocked	Bank accountChildren's personal dataMobile number	Critical	10 days ago
“Draft a summary of health claim patterns by district” Aggregate over special-category data; secondary authorisation required.	Departmental AI Assistant	step up		High	14 days ago
“Extract all names and IDs from the attached beneficiary list” Bulk extraction from an attachment classified as high-risk personal data.	ChatGPT (consumer)	blocked	Bank accountMobile numberChildren's personal data	Critical	14 days ago
“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.	GitHub Copilot	blocked	Bank accountMobile numberGrievance history	Critical	yesterday
“What is the bank account on file for Family ID HR-08-••••4471?” Direct retrieval of a high-risk identifier category.	Local translation model	blocked	Grievance historyMobile number	Critical	2 days ago
“List students in the scholarship shortlist with their dates of birth” Children's personal data — blocked by the children's data guardrail.	Grievance triage agent (pilot)	blocked	Mobile numberGrievance historyChildren's personal data	Critical	15 days ago
“Give me the phone numbers of everyone who complained about ration cards this month” Enumeration pattern over classified contact data.	Microsoft 365 Copilot	blocked	Children's personal dataBank account	Critical	14 days ago
“Summarise this case file” Output contained identifier and bank account values; masked before display.	Gemini (consumer)	redacted	Government identifierBank account	Medium	3 days ago
“Who lives at the address on application SCH-2026-3312?” Inference-shaped prompt targeting a restricted category.	Embedded summarisation in the case tool	blocked	Mobile number	Critical	yesterday
“Show me records for citizens who withdrew consent last week” Consent state is withdrawn; retrieval denied and logged.	ChatGPT (consumer)	blocked	Children's personal dataGovernment identifier	Critical	6 days ago
“Draft a summary of health claim patterns by district” Aggregate over special-category data; secondary authorisation required.	GitHub Copilot	step up		High	6 days ago
“Extract all names and IDs from the attached beneficiary list” Bulk extraction from an attachment classified as high-risk personal data.	Local translation model	blocked	Grievance historyBank accountChildren's personal data	Critical	12 days ago
“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.	Scheme eligibility scoring model	blocked	Bank account	Critical	5 days ago
“What is the bank account on file for Family ID HR-08-••••4471?” Direct retrieval of a high-risk identifier category.	Grievance triage agent (pilot)	blocked	Grievance history	Critical	14 days ago
“List students in the scholarship shortlist with their dates of birth” Children's personal data — blocked by the children's data guardrail.	Microsoft 365 Copilot	blocked	Grievance history	Critical	4 days ago
“Give me the phone numbers of everyone who complained about ration cards this month” Enumeration pattern over classified contact data.	Embedded summarisation in the case tool	blocked	Grievance history	Critical	15 days ago
“Summarise this case file” Output contained identifier and bank account values; masked before display.	Departmental AI Assistant	redacted	Government identifierBank account	Medium	5 days ago
“Who lives at the address on application SCH-2026-3312?” Inference-shaped prompt targeting a restricted category.	ChatGPT (consumer)	blocked	Government identifier	Critical	11 days ago
“Show me records for citizens who withdrew consent last week” Consent state is withdrawn; retrieval denied and logged.	Local translation model	blocked	Government identifier	Critical	12 days ago

1–25 of 66

1 / 3

Evidences:S.8(4)S.9 R.6Output enforcement is the control that catches inference — where a model assembles a disclosing answer from data that individually looked harmless.