AI Processing Guard

SL-09

AI systems that process, infer from or generate outputs using personal data are subject to the same safeguarding, breach and accountability obligations as any other processing. This module inspects prompts and outputs, validates consent and purpose before retrieval, and finds the AI nobody told you about.

DischargesS.4 S.6 S.8(4)S.9 S.16 R.6

Shadow AI Guardrails

3 unsanctioned AI services in use

ChatGPT (consumer), Gemini (consumer), Claude (consumer) — all outside the department's boundary, most with data residency outside India. Every prompt containing personal data is blocked, but the attempt itself indicates a workflow gap people are routing around.

Prompts (30 days)

79.4 K

Blocked

3.77 K

5% of traffic

Outside India

Touching children's data

Prompt volume and enforcement

AI usage grows faster than governance does, which is why this is a monitoring problem before it is a policy problem.

Systems by type

Shadow AI shown in red.

Shadow AI3
Hosted AI platform1
Enterprise copilot1
Analytics / ML model1
Agentic framework1
AI development tool1
Embedded AI feature1
Custom LLM1

Traffic by system

Blocked proportion is the honest measure of how much personal data people are trying to push through each tool.

All systems

Recent enforcement

Prompts and outputs the guardrails acted on.

Full feed

blockedGitHub Copilot · Sunita Pandeyyesterday
“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.
blockedEmbedded summarisation in the case tool · Pallavi Reddyyesterday
“Who lives at the address on application SCH-2026-3312?”
Inference-shaped prompt targeting a restricted category.
blockedEmbedded summarisation in the case tool · Yash Patelyesterday
“List students in the scholarship shortlist with their dates of birth”
Children's personal data — blocked by the children's data guardrail.
blockedLocal translation model · Vikram Sethiyesterday
“Extract all names and IDs from the attached beneficiary list”
Bulk extraction from an attachment classified as high-risk personal data.
blockedLocal translation model · Pallavi Reddy2 days ago
“What is the bank account on file for Family ID HR-08-••••4471?”
Direct retrieval of a high-risk identifier category.
blockedChatGPT (consumer) · Yash Patel2 days ago
“Show me records for citizens who withdrew consent last week”
Consent state is withdrawn; retrieval denied and logged.

Active guardrails

Each consumes Discovery classification so decisions are made on actual data sensitivity rather than generic rules.

Configure

Block prompts requesting classified personal dataPromptsPurposeMass extractionblock284Active
Redact personal data in AI outputsOutputsChildrenredact1842Active
Enforce consent and purpose before AI retrievalPromptsOutputsConsentPurposeChildrenblock96Active
Block all unsanctioned AI servicesPromptsChildrenMass extractionblock2900Active
Developer tooling — source onlyPromptsChildrenwarn18Active

2 AI systems can reach children's personal data

Departmental AI Assistant, ChatGPT (consumer). Section 9 prohibits tracking and behavioural monitoring of children entirely, and an AI system that can infer attributes is doing something close to both. Exclude children's data classes from indexing rather than relying on output filtering alone.

AI system inventoryEvery generative model, copilot, agentic framework, embedded feature and analytics model that can reach personal data.Prompt inspectionWhat was asked, what was blocked and why — including mass-extraction and enumeration patterns.Output controlsResponses analysed for personal data and masked, blocked or held for secondary authorisation.Interaction auditComplete input and output traceability — the evidence that preventive controls existed and acted.

Related