Guardrail Policies

SL-09

Constraints applied to every AI interaction, driven by what Discovery classified the data as rather than by generic keyword rules. That is what allows a guardrail to permit a legitimate case-drafting prompt while blocking a district-wide contact extraction.

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

Progressive enforcement, not a binary switch

Guardrails escalate — warn, then restrict, then suspend — and tighten automatically when an identity is already under investigation. Jumping straight to blocking on a moderate-confidence signal creates the kind of friction that gets controls disabled.

Guardrails

Active

Consent-validating

Deny after withdrawal

Interventions (30 days)

5140

Configured guardrails

Block prompts requesting classified personal dataon violation: blockProgressive

Inspects prompts for attempts to retrieve data classified as personal or high-risk, and for mass-extraction or enumeration patterns.

284

30-day hits

Checks performed

Inspects prompts
Reads the request before any retrieval happens.
Inspects outputs
Analyses the response for personal data before display.
Validates consent
Blocks retrieval where consent has been withdrawn or has expired.
Validates purpose
Checks the requested data is within the purpose it was collected for.
Blocks children's data
Prevents any retrieval of data classified as a child's personal data.
Detects mass extraction
Recognises enumeration and bulk-retrieval patterns in the prompt itself.

Applies toDepartmental AI AssistantM365 CopilotGrievance triage agent

S.4 S.8(4)R.6Updated 19 Jul 2026

Redact personal data in AI outputson violation: redact

Analyses every response for personal or sensitive content and masks identifier categories before display.

1842

30-day hits

Checks performed

Inspects prompts
Reads the request before any retrieval happens.
Inspects outputs
Analyses the response for personal data before display.
Validates consent
Blocks retrieval where consent has been withdrawn or has expired.
Validates purpose
Checks the requested data is within the purpose it was collected for.
Blocks children's data
Prevents any retrieval of data classified as a child's personal data.
Detects mass extraction
Recognises enumeration and bulk-retrieval patterns in the prompt itself.

Applies toDepartmental AI AssistantM365 CopilotEmbedded summarisation

S.8(4)S.9Updated 26 Jul 2026

Enforce consent and purpose before AI retrievalon violation: block

Validates that the data a prompt would retrieve is within the declared purpose and that consent is active. Blocks where consent is withdrawn or expired.

30-day hits

Checks performed

Inspects prompts
Reads the request before any retrieval happens.
Inspects outputs
Analyses the response for personal data before display.
Validates consent
Blocks retrieval where consent has been withdrawn or has expired.
Validates purpose
Checks the requested data is within the purpose it was collected for.
Blocks children's data
Prevents any retrieval of data classified as a child's personal data.
Detects mass extraction
Recognises enumeration and bulk-retrieval patterns in the prompt itself.

Applies toDepartmental AI AssistantGrievance triage agent

S.6 S.4 S.9Updated 31 Jul 2026

Block all unsanctioned AI serviceson violation: block

Any personal data destined for an AI service outside the sanctioned catalogue is blocked at the gateway and logged.

2900

30-day hits

Checks performed

Inspects prompts
Reads the request before any retrieval happens.
Inspects outputs
Analyses the response for personal data before display.
Validates consent
Blocks retrieval where consent has been withdrawn or has expired.
Validates purpose
Checks the requested data is within the purpose it was collected for.
Blocks children's data
Prevents any retrieval of data classified as a child's personal data.
Detects mass extraction
Recognises enumeration and bulk-retrieval patterns in the prompt itself.

Applies toChatGPT (consumer)Gemini (consumer)Claude (consumer)

S.8(4)S.16 R.6Updated 28 Jul 2026

Developer tooling — source onlyon violation: warnProgressive

Permits code context but blocks any prompt containing content classified as personal data.

30-day hits

Checks performed

Inspects prompts
Reads the request before any retrieval happens.
Inspects outputs
Analyses the response for personal data before display.
Validates consent
Blocks retrieval where consent has been withdrawn or has expired.
Validates purpose
Checks the requested data is within the purpose it was collected for.
Blocks children's data
Prevents any retrieval of data classified as a child's personal data.
Detects mass extraction
Recognises enumeration and bulk-retrieval patterns in the prompt itself.

Applies toGitHub Copilot

R.6Updated 03 Jul 2026

The dependency worth understanding

None of these guardrails work without Discovery. “Block children's data” is only meaningful if something has already classified which columns and files contain it; “deny after withdrawal” is only meaningful if consent state is bound to data elements. An AI guard deployed without a classification layer underneath it degrades to keyword filtering, which fails in both directions.