feat: Singapore guardrail policies (PDPA + MAS AI Risk Management)

[ron-zhong]

Summary

• Add Singapore PDPA based guardrail policy template with regex + conditional keyword guardrails
• Add MAS AI Risk Management Guidelines based guardrail policy template for financial institutions
• Add Singapore-specific regex patterns (NRIC/FIN, phone, postal code, passport, UEN, bank account)
• Add policy template entries to policy registry and backup registry
• Add focused tests for Singapore PDPA and MAS guardrails

What changed

PDPA (Personal Data Protection Act)

• Added 6 regex patterns in the content filter patterns registry:
  • sg_nric — Singapore NRIC/FIN format detection ([STFGM] + 7 digits + checksum letter)
  • sg_phone — Singapore phone numbers (supports +65, 0065, 65 prefixes)
  • sg_postal_code — Singapore 6-digit postal code detection (contextual)
  • passport_singapore — Singapore passport numbers (E/K + 7 digits, contextual)
  • sg_uen — Singapore Unique Entity Number detection (all 3 common formats)
  • sg_bank_account — Singapore-style bank account numbers (dash format, contextual)
• Added 5 YAML policy templates:
  • sg_pdpa_personal_identifiers
  • sg_pdpa_sensitive_data
  • sg_pdpa_do_not_call
  • sg_pdpa_data_transfer
  • sg_pdpa_profiling_automated_decisions
• Added pdpa-singapore policy template entry to policy_templates.json and litellm/policy_templates_backup.json
• Added tests:
  • tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_sg_patterns.py
  • tests/guardrails_tests/test_sg_pdpa_guardrails.py

MAS (Monetary Authority of Singapore) AI Risk Management Guidelines

• Added 5 YAML policy templates:
  • sg_mas_fairness_bias
  • sg_mas_transparency_explainability
  • sg_mas_human_oversight
  • sg_mas_data_governance
  • sg_mas_model_security
• Added mas-ai-risk-management policy template entry to policy_templates.json and litellm/policy_templates_backup.json
• Added tests:
  • tests/guardrails_tests/test_sg_mas_ai_guardrails.py

Notes

• Followed existing policy template and guardrail test patterns from prior regulatory templates.
• Kept changes data-driven (YAML + JSON + tests), no core runtime logic changes.

Reference:

• https://www.pdpc.gov.sg/overview-of-pdpa/the-legislation/personal-data-protection-act
• https://www.pdpc.gov.sg/help-and-resources/2020/01/model-ai-governance-framework
• https://www.mas.gov.sg/publications/consultations/2025/consultation-paper-on-guidelines-on-artificial-intelligence-risk-management

Related PRs:

• feat: split EU AI Act Article 5 into 5 dedicated sub-guardrails #21453
• feat: EU AI Act Article 5 policy template for prohibited practices detection #21342

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 23, 2026 7:36pm

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

Adds two Singapore-specific guardrail policy templates — PDPA (Personal Data Protection Act) and MAS AI Risk Management (for financial institutions) — along with 6 new regex patterns for Singapore PII detection (NRIC/FIN, phone, postal code, passport, UEN, bank account).

• PDPA template includes 9 guardrails: 4 regex-based PII masking guardrails and 5 keyword-based conditional matching guardrails covering consent (s.13), sensitive data (Advisory Guidelines), DNC Registry (Part IX), overseas data transfers (s.26), and automated profiling (Model AI Governance Framework).
• MAS AI Risk Management template includes 5 keyword-based conditional matching guardrails covering fairness/bias, transparency/explainability, human oversight, data governance, and model security.
• All changes are data-driven (YAML + JSON) with no core runtime logic modifications.
• Tests are comprehensive (156 regex pattern tests + ~130 conditional keyword matching test cases across PDPA and MAS), follow existing patterns from prior regulatory templates (EU AI Act), and make no network calls.
• Minor inconsistency: policy_templates.json and litellm/policy_templates_backup.json have divergent MAS guardrail descriptions (em dash — vs hyphen -) and different templateData.description wording for the MAS template. This divergence already exists for other templates in the backup file, so it may be by design.

Confidence Score: 4/5

• This PR is safe to merge — it adds only data-driven configuration (YAML, JSON) and tests with no runtime logic changes.
• Score of 4 reflects: all changes are data-driven (no runtime code modifications), comprehensive test coverage for both regex patterns and keyword matching, follows established patterns from prior regulatory templates. Minor deduction for the inconsistency between policy_templates.json and the backup file, but this is a pre-existing pattern in the codebase.
• The litellm/policy_templates_backup.json has description inconsistencies with policy_templates.json for the MAS guardrail entries (em dash vs hyphen, different templateData.description wording).

Important Files Changed

Filename	Overview
litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/patterns.json	Adds 6 Singapore PII regex patterns (NRIC, phone, postal code, passport, UEN, bank account). Patterns are well-designed with keyword_pattern context filters where appropriate. The sg_uen pattern now includes keyword_pattern (addressing prior review feedback). Minor: file still lacks trailing newline (pre-existing).
policy_templates.json	Adds two new policy template entries: pdpa-singapore (9 guardrails covering PII, sensitive data, DNC, data transfer, profiling) and mas-ai-risk-management (5 guardrails covering fairness, transparency, oversight, data governance, model security). Descriptions are consistent within the file. Minor inconsistency with backup file (em dash vs hyphen and templateData.description wording).
litellm/policy_templates_backup.json	Backup copy of policy templates with same two new entries. Contains description inconsistencies with the primary file: uses hyphens where primary uses em dashes, and the MAS templateData.description wording differs. This file is a backup and already diverges from primary in other ways (has example_sentences, extra templates).
litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_pdpa_personal_identifiers.yaml	PDPA s.13 policy template for blocking unauthorized collection of Singapore personal identifiers. Well-structured with identifier_words, additional_block_words, always_block_keywords, and exceptions. Follows existing patterns.
litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_fairness_bias.yaml	MAS fairness and bias policy template. Blocks discriminatory AI practices in financial services based on protected attributes, with Singapore-specific ethnic groups (CMIO). Well-structured exceptions for fairness audits and research.
litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_model_security.yaml	MAS model security policy template. Blocks adversarial attacks, data poisoning, model inversion, and exfiltration targeting financial AI systems. Appropriate exceptions for red-teaming and security research.
tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_sg_patterns.py	Unit tests for the 6 Singapore PII regex patterns. Tests valid/invalid format detection for NRIC, phone, postal code, passport, UEN, and bank account. Previous review feedback about case-insensitive test has been addressed (line 36 now expects match). No network calls — pure regex testing.
tests/guardrails_tests/test_sg_pdpa_guardrails.py	Integration tests for 5 PDPA sub-guardrails with parametrized test cases covering always-block, conditional matching, exceptions, and no-match scenarios. No network calls. Follows existing patterns from EU AI Act tests.
tests/guardrails_tests/test_sg_mas_ai_guardrails.py	Integration tests for 5 MAS sub-guardrails with parametrized test cases covering always-block, conditional matching, exceptions, and no-match scenarios. No network calls. Follows existing patterns from other guardrail tests.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    subgraph PDPA["PDPA Singapore Policy Template"]
        PII["pdpa-sg-pii-identifiers\n(regex: NRIC, passport)"]
        CONTACT["pdpa-sg-contact-information\n(regex: phone, postal, email)"]
        FIN["pdpa-sg-financial-data\n(regex: bank account, credit card)"]
        BIZ["pdpa-sg-business-identifiers\n(regex: UEN)"]
        PI["pdpa-sg-personal-identifiers\n(keyword: s.13 Consent)"]
        SD["pdpa-sg-sensitive-data\n(keyword: Advisory Guidelines)"]
        DNC["pdpa-sg-do-not-call\n(keyword: Part IX DNC)"]
        DT["pdpa-sg-data-transfer\n(keyword: s.26 Transfer)"]
        PROF["pdpa-sg-profiling-automated-decisions\n(keyword: AI Governance)"]
    end

    subgraph MAS["MAS AI Risk Management Policy Template"]
        FB["mas-sg-fairness-bias\n(keyword)"]
        TE["mas-sg-transparency-explainability\n(keyword)"]
        HO["mas-sg-human-oversight\n(keyword)"]
        DG["mas-sg-data-governance\n(keyword)"]
        MS["mas-sg-model-security\n(keyword)"]
    end

    subgraph Patterns["patterns.json — New Regex Patterns"]
        NRIC["sg_nric"]
        PHONE["sg_phone"]
        POSTAL["sg_postal_code"]
        PASS["passport_singapore"]
        UEN["sg_uen"]
        BANK["sg_bank_account"]
    end

    PII --> NRIC
    PII --> PASS
    CONTACT --> PHONE
    CONTACT --> POSTAL
    FIN --> BANK
    BIZ --> UEN

    INPUT["User Input"] --> PDPA
    INPUT --> MAS
    PDPA -->|MASK/BLOCK| OUTPUT["Filtered Output"]
    MAS -->|BLOCK| OUTPUT

Loading

_{Last reviewed commit: fe00f46}

[greptile-apps]

_{16 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[ron-zhong]

Addressed both points from the review:

1. Failing lowercase NRIC test

• Updated tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_sg_patterns.py to align with re.IGNORECASE behavior in get_compiled_pattern():
  • test_lowercase_letter_prefix_rejected -> test_lowercase_letter_prefix_detected_case_insensitive
  • assertion updated to expect a match for s1234567A

2. sg_uen false-positive risk (missing context)

• Updated litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/patterns.json for sg_uen with keyword gating:
  • "keyword_pattern": "\\b(?:UEN|company\\s*registration|business\\s*registration|ACRA|entity\\s*number|ROC|ROS)\\b"
  • "allow_word_numbers": false

Also added a case-insensitive UEN test (12345678a) to reflect runtime matching behavior.

[ron-zhong]

Files updated in this follow-up commit:

• policy_templates.json
• litellm/policy_templates_backup.json
• litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_fairness_bias.yaml
• litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_human_oversight.yaml
• litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_transparency_explainability.yaml

Commit: 6db71ad

[ron-zhong]

Resolved conflicts have been committed and pushed in caf9baa459.

Ready for re-review.

cc @ishaan-jaff @krrishdholakia

[krrishdholakia]

@greptile please re-review this

[greptile-apps]

_{16 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[ron-zhong]

Addressed this review feedback in the latest push:

• Synced MAS guardrail descriptions between policy_templates.json and litellm/policy_templates_backup.json (using em dashes consistently).

Both registry entries are now aligned.

[ron-zhong]

Hi @krrishdholakia, will there be an image build that we can download (docker pull) for an internal preview before merging our contributions to the release branch? Thank you.

(Updated) I've found the build here, thanks.

• https://github.com/BerriAI/litellm/releases/tag/litellmv1.81.15.presidio.dev

[shin-bot-litellm]

Hey @ron-zhong! Great PR on the Singapore guardrail policies — really solid work adding PDPA and MAS AI Risk Management support.

Your background as Solution Architect / DevOps Engineer at MAS is impressive. We're hiring engineers at LiteLLM and thought you might be interested!

If you're open to chatting, here's Ishaan's (our CTO) Calendly: https://calendly.com/ishaan-berri2/litellm-founding-engineer-round-1-interview-clone

Either way, thanks for the contribution!

[ron-zhong]

Relevant PRs:

• refactor(policies): prefix SG guardrail policy IDs with country code #21974
• Fix replicate to foreign server without PDPA #21978