* Add MCP_SECURITY enum to SupportedGuardrailIntegrations
* Add MCP security guardrail initializer
* Add MCPSecurityGuardrail implementation
* Add MCP Security policy template
* Add Type filter to policy templates UI
* Add unit tests for MCP security guardrail
* fix(lint): remove unused Dict import from mcp_security_guardrail
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Add French language support for EU AI Act Article 5 guardrail (#21427)
* Add French language support for EU AI Act Article 5 template
- Create eu_ai_act_article5_fr.yaml with comprehensive French keywords
- Includes identifier words: concevoir, créer, développer, noter, classer, etc.
- Includes block words: crédit social, comportement social, émotion des employés, etc.
- Includes always-block keywords for explicit prohibited practices
- Includes exceptions for research, compliance, and legitimate use cases
- Catches circumvention attempts with phrase variations
* Add comprehensive tests for French EU AI Act guardrail
- Test 3 critical scenarios: blocked query, circumvention attempt, safe query
- Test edge cases: case-insensitive, mixed language, research exceptions
- All 7 tests passing
- Validates both blocking and allowing behavior
* Fix content filter to support conditional matching without inherit_from
- Enable conditional matching when identifier_words + additional_block_words are present
- Previously required inherit_from, but EU AI Act templates are self-contained
- Fixes Greptile feedback: conditional matching now works as documented
* Add pure conditional matching test for French guardrail
- Test identifier + block word combinations not in always_block_keywords
- Verifies conditional matching works independently
- Addresses Greptile feedback about test coverage gap
* Fix exception word bypass risk in French template
- Replace short words (film, jeu, juste) with context-specific phrases
- Prevents substring matching bypasses (e.g., enjeu matching jeu)
- Add tests for bypass prevention and legitimate game context
- Addresses Greptile security feedback
* Make conditional match assertion more robust
- Use getattr to safely access exception detail field
- Check if detail is dict before calling .get()
- Addresses Greptile feedback about brittle string assertion
* Add French EU AI Act Article 5 policy template to registry
- Add eu-ai-act-article5-fr template for French language support
- Includes French description and guardrail info
- Matches structure of English template
* Address greptile review feedback (greploop iteration 1)
- Use status_code=400 instead of 403 to match guardrail logging convention
- Use prefix stripping instead of split('/')[-1] for robust server name extraction
* remove French EU AI Act template from policy_templates.json
---------
Co-authored-by: Julio Quinteros Pro <jquinter@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Use importlib.import_module + reload uniformly in both code paths
to ensure fresh module state regardless of whether litellm was
previously in sys.modules. This fixes the inconsistency where the
"not in sys.modules" branch didn't reload the module.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- test_pillar_guardrails.py: Fix fixture to properly update module-level
litellm reference using global keyword and assignment from reload
- test_anthropic_experimental_pass_through_messages_handler.py: Add missing
assert keywords to kwargs comparison statements (lines 36, 60-62)
- test_proxy_server.py: Replace silent pytest.skip with explicit assertion
to catch router initialization regressions
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Add 6 new EU PII patterns for GDPR compliance
- fr_nir: French Social Security Number (NIR/INSEE) with validation
- eu_iban_enhanced: Enhanced IBAN detection with specific format
- fr_phone: French phone numbers (+33, 0033, 0 formats)
- eu_vat: EU VAT identification numbers (all 27 member states)
- eu_passport_generic: Generic EU passport format
- fr_postal_code: French postal codes with contextual keywords
* Add GDPR Art. 32 EU PII Protection policy template
- Comprehensive GDPR Article 32 compliance policy
- 4 guardrail groups: National IDs, Financial, Contact Info, Business IDs
- Masks French NIR/INSEE, EU IBANs, French phones, EU VAT numbers
- Includes EU passport numbers and email addresses
- Medium complexity template with indigo icon
* Add comprehensive tests for EU PII patterns
- Test French NIR validation (sex digit, month range)
- Test enhanced IBAN detection (French, German)
- Test French phone number formats
- Test EU VAT numbers
- Test generic EU passport format
- Test French postal code pattern
* Add EU pattern loading and category validation tests
- Verify all 6 EU PII patterns are loaded correctly
- Verify patterns are categorized as 'EU PII Patterns'
- Ensure pattern loading consistency
* Add end-to-end tests for GDPR policy template
- 4 tests for PII that should be masked (NIR, IBAN, phone, VAT)
- 4 tests for text that should pass through (invalid patterns, no PII)
- 1 bonus test for multiple PII types in same message
- All tests verify correct masking behavior
* Add region field to policy templates
- Added region field to all 6 templates (EU, AU, Global)
- Updated both main and backup JSON files
- Enables region-based filtering in UI
* Add region filter to policy templates UI
- Added Radio.Group filter for regions (All, AU, EU, Global)
- Efficient filtering with useMemo hooks
- Clean button-based UI matching existing design
- Defaults missing regions to Global
* Apply suggestion from @greptile-apps[bot]
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* Address Greptile review: add contextual guards and negative tests
- Added keyword_pattern to eu_vat (VAT, tax number, fiscal code, etc.)
- Added keyword_pattern to eu_passport_generic (passport, travel document, etc.)
- Added 3 negative unit tests for false positive prevention
- Added 2 E2E tests verifying no masking without keyword context
- All patterns now require contextual keywords to prevent false positives
* address greptile review feedback (greploop iteration 1)
- Remove unused HTTPException import from test file
- Add keyword_pattern to eu_vat for contextual VAT matching
- Add allow_word_numbers: false to eu_passport_generic
- Add negative test cases for EU VAT false positives
- All 5 Greptile comments addressed
* Address Greptile feedback: fix patterns and sync backup
- Fix fr_phone pattern: use negative lookbehind (?<!\d) to prevent false matches in longer digit strings
- Add keyword_pattern to eu_passport_generic to reduce false positives on version strings/SKUs
- Sync policy_templates_backup.json with main file (add GDPR template)
- Add keyword_pattern to eu_vat (was auto-added by formatter)
All pattern tests passing
* address greptile review feedback (greploop iteration 2)
- Update test to document that eu_vat raw pattern is intentionally broad
- Test verifies pattern DOES match common words (by design)
- Documents that keyword_pattern guard prevents false positives in production
- Addresses Greptile's false positive risk concern
* address greptile review feedback (greploop iteration 3)
- Fix test_eu_vat_masked: change gap from 2 words to 1 word (VAT number: FR...)
- This ensures keyword matching works within MAX_KEYWORD_VALUE_GAP_WORDS=1 limit
- fr_phone pattern already works correctly (verified with tests)
- test_pattern_requires_keyword_context already updated in iteration 2
Addresses final issues from Greptile 2/5 review
* fix: remove country-specific passport patterns from GDPR template
- Remove passport_france, passport_germany, passport_netherlands from template
- These patterns lack keyword guards and cause false positives
- Only eu_passport_generic remains (has keyword_pattern guard)
- Sync policy_templates_backup.json with main file
- Update test setup to match template
All 11 E2E tests now passing ✅
* Update tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_gdpr_policy_e2e.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* Update litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
---------
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
The setup_and_teardown fixture was failing with "ImportError: module
litellm not in sys.modules" during parallel test execution. This occurs
because another worker might have removed/modified litellm from
sys.modules before this test tries to reload it.
Fix: Check if litellm is in sys.modules before attempting reload.
* fix(content_filter.py): fix filter on toxic keywords
* feat: improve toxic/abusive language detection
* fix: additional improvements to nsfw filters
* feat: more improvements to nsfw filter
* feat(content_filter.json): add new australia specific nsfw content filter
ensure complete coverage for australia nsfw
* fix: cleanup policy templates
* fix(index.tsx): alert notice
* fix(index.tsx): add disclaimer notice
* feat(harmful_child_safety.yaml): new child safety content filter
ensure we catch inappropriate, child-specific content
* feat(policy_templates.json): add child safety and self harm filters
* fix(content_filter.py): improve racial bias filter to use a similar identifier + block word pattern and cover a wider range of ethnicities
* feat(policy_templates.json): add racial bias to nsfw policy template
* feat: add json content viewer
* fix: empty guardrails/policies arrays should not trigger enterprise license check (#20304)
The UI sends empty arrays for enterprise-only fields (guardrails, policies,
logging) even when the user has not configured these features. The backend
`is not None` check treated `[]` as a truthy intent to use the feature,
falsely requiring an enterprise license for basic team operations.
Backend: Add `and updated_kv[field] != [] and updated_kv[field] != {}`
guards in `_update_metadata_fields` so empty collections are skipped.
UI: Conditionally omit guardrails, logging, and policies from the
payload when empty instead of defaulting to `[]`.
Fixes#20304
* fix: allow clearing fields with empty collections while skipping enterprise check
Address PR review feedback:
1. Move the empty-collection guard into _update_metadata_field (singular)
so that empty lists/dicts skip only the premium license check but still
get written into metadata. This lets users intentionally clear a
previously-set field (e.g. guardrails: []) without being blocked, while
the UI's default empty arrays still don't trigger a false enterprise
error.
2. Remove sys.path hack from test file; use standard imports that work
with pytest discovery.
3. Add tests verifying that empty collections are moved into metadata
(field clearing works) even though they bypass the premium check.
Fixes#20304
* fix(proxy): add regression tests for #20441 - ensure <script> tags in LLM messages are not blocked
The 403 Forbidden error when sending messages containing `<script>` is caused
by external WAF/reverse proxy infrastructure (confirmed by the standard nginx
HTML 403 response format), not by LiteLLM's own content filtering. However,
these regression tests ensure that:
1. The content filter guardrail's built-in patterns do not match HTML tags
2. Messages containing <script> and other HTML tags pass through the content
filter unchanged when no explicit HTML-blocking rules are configured
3. The HTTP request body parser correctly handles JSON payloads containing
HTML content without modification
These tests guard against accidentally introducing HTML/XSS filtering that
would break legitimate LLM API usage (e.g., discussing HTML/JavaScript code).
Closes#20441
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add opus 4.5 and 4.6 to use outout_format param
* generate poetry lock with 2.3.2 poetry
* restore poetry lock
* e2e tests, key delete, update tpm rpm, and regenerate
* Split e2e ui testing for browser
* new login with sso button in login page
* option to hide usage indicator
* fix(cloudzero): update CBF field mappings per LIT-1907 (#20906)
* fix(cloudzero): update CBF field mappings per LIT-1907
Phase 1 field updates for CloudZero integration:
ADD/UPDATE:
- resource/account: Send concat(api_key_alias, '|', api_key_prefix)
- resource/service: Send model_group instead of service_type
- resource/usage_family: Send provider instead of hardcoded 'llm-usage'
- action/operation: NEW - Send team_id
- resource/id: Send model name instead of CZRN
- resource/tag:organization_alias: Add if exists
- resource/tag:project_alias: Add if exists
- resource/tag:user_alias: Add if exists
REMOVE:
- resource/tag:total_tokens: Removed
- resource/tag:team_id: Removed (team_id now in action/operation)
Fixes LIT-1907
* Update litellm/integrations/cloudzero/transform.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* fix: define api_key_alias variable, update CBFRecord docstring
- Fix F821 lint error: api_key_alias was used but not defined
- Update CBFRecord docstring to reflect LIT-1907 field mappings
- Remove unused Optional import
---------
Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* Add banner notifying of breaking change
* Add semgrep & Fix OOMs (#20912)
* [Feat] Policies - Allow connecting Policies to Tags, Simulating Policies, Viewing how many keys, teams it applies on (#20904)
* init schema with TAGS
* ui: add policy test
* resolvePoliciesCall
* add_policy_sources_to_metadata + headers
* types Policy
* preview Impact
* def _describe_match_reason(
* match based on TAGs
* TestTagBasedAttachments
* test fixes
* add policy_resolve_router
* add_guardrails_from_policy_engine
* TestMatchAttribution
* refactor
* fix
* fix: address Greptile review feedback on policy resolve endpoints
- Track unnamed keys/teams as separate counts instead of inflating
affected_keys_count with duplicate "(unnamed key)" placeholders.
Added unnamed_keys_count and unnamed_teams_count to response.
- Push alias pattern matching to DB via _build_alias_where() which
converts exact patterns to Prisma "in" and suffix wildcards to
"startsWith" filters.
- Gate sync_policies_from_db/sync_attachments_from_db behind
force_sync query param (default false) to avoid 2 DB round-trips
on every /policies/resolve request.
- Remove worktree-only conftest.py that cleared sys.modules at import
time — no longer needed since code moved to main repo.
- Rename MAX_ESTIMATE_IMPACT_ROWS → MAX_POLICY_ESTIMATE_IMPACT_ROWS.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: eliminate duplicate DB queries and fix header delimiter ambiguity
- Fetch teams table once in estimate_attachment_impact and reuse for
both tag-based and alias-based lookups (was querying teams twice when
both tag_patterns and team_patterns were provided).
- Convert tag/team filter functions from async DB queries to sync
filters that operate on pre-fetched data (_filter_keys_by_tags,
_filter_teams_by_tags).
- Fix comma ambiguity in x-litellm-policy-sources header: use '; '
as entry delimiter since matched_via values can contain commas.
- Use '+' as the within-value separator in matched_via reason strings
(e.g. "tag:healthcare+team:health-team") to avoid conflict with
header delimiters.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Update litellm/proxy/policy_engine/policy_resolve_endpoints.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* fix: type error & better error handling (#20689)
* [Docs] Add docs guide for using policies (#20914)
* init schema with TAGS
* ui: add policy test
* resolvePoliciesCall
* add_policy_sources_to_metadata + headers
* types Policy
* preview Impact
* def _describe_match_reason(
* match based on TAGs
* TestTagBasedAttachments
* test fixes
* add policy_resolve_router
* add_guardrails_from_policy_engine
* TestMatchAttribution
* refactor
* fix
* fix: address Greptile review feedback on policy resolve endpoints
- Track unnamed keys/teams as separate counts instead of inflating
affected_keys_count with duplicate "(unnamed key)" placeholders.
Added unnamed_keys_count and unnamed_teams_count to response.
- Push alias pattern matching to DB via _build_alias_where() which
converts exact patterns to Prisma "in" and suffix wildcards to
"startsWith" filters.
- Gate sync_policies_from_db/sync_attachments_from_db behind
force_sync query param (default false) to avoid 2 DB round-trips
on every /policies/resolve request.
- Remove worktree-only conftest.py that cleared sys.modules at import
time — no longer needed since code moved to main repo.
- Rename MAX_ESTIMATE_IMPACT_ROWS → MAX_POLICY_ESTIMATE_IMPACT_ROWS.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: eliminate duplicate DB queries and fix header delimiter ambiguity
- Fetch teams table once in estimate_attachment_impact and reuse for
both tag-based and alias-based lookups (was querying teams twice when
both tag_patterns and team_patterns were provided).
- Convert tag/team filter functions from async DB queries to sync
filters that operate on pre-fetched data (_filter_keys_by_tags,
_filter_teams_by_tags).
- Fix comma ambiguity in x-litellm-policy-sources header: use '; '
as entry delimiter since matched_via values can contain commas.
- Use '+' as the within-value separator in matched_via reason strings
(e.g. "tag:healthcare+team:health-team") to avoid conflict with
header delimiters.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* docs v1 guide with UI imgs
* docs fix
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add dashscope/qwen3-max model with tiered pricing (#20919)
Add support for Alibaba Cloud's Qwen3-Max model with:
- 258K input tokens, 65K output tokens
- Tiered pricing based on context window usage (0-32K, 32K-128K, 128K-252K)
- Function calling and tool choice support
- Reasoning capabilities enabled
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
* fix linting
* docs: add Greptile review requirement to PR template (#20762)
* fix(azure): preserve content_policy_violation error details from Azure OpenAI
Closes#20811
Azure OpenAI returns rich error payloads for content policy violations
(inner_error with ResponsibleAIPolicyViolation, content_filter_results,
revised_prompt). Previously these details were lost when:
1. The top-level error code was not "content_policy_violation" but the
inner_error.code was "ResponsibleAIPolicyViolation" -- the structured
check only examined the top-level code.
2. The DALL-E image generation polling path stringified the error JSON
into the message field instead of setting the structured body, making
it impossible for exception_type() to extract error details.
3. The string-based fallback detector used "invalid_request_error" as a
content-policy indicator, which is too broad and could misclassify
regular bad-request errors.
Changes:
- exception_mapping_utils.py: Check inner_error.code for
ResponsibleAIPolicyViolation when top-level code is not
content_policy_violation. Replace overly broad "invalid_request_error"
string match with specific Azure safety-system messages.
- azure.py: Set structured body on AzureOpenAIError in both async and
sync DALL-E polling paths so exception_type() can inspect error details.
- test_azure_exception_mapping.py: Add regression tests covering the
exact error payloads from issue #20811.
- Fix pre-existing lint: duplicate PerplexityResponsesConfig dict key,
unused RouteChecks top-level import.
---------
Co-authored-by: Kelvin Tran <kelvin-tran@users.noreply.github.com>
Co-authored-by: yuneng-jiang <yuneng.jiang@gmail.com>
Co-authored-by: shin-bot-litellm <shin-bot-litellm@berri.ai>
Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Co-authored-by: Alexsander Hamir <alexsanderhamirgomesbaptista@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Harshit Jain <48647625+Harshit28j@users.noreply.github.com>
Co-authored-by: ken <122603020@qq.com>
Co-authored-by: Sameer Kankute <sameer@berri.ai>
* Generic Guardrails: Forward request headers + litellm_version to generic guardrail API
* Generic Guardrail: Change the request headers addition to be with allowlist instead denylist
The unified guardrail's streaming iterator hook processes every Nth
chunk (sampling_rate, default 5). On each sampled chunk it calls
process_output_streaming_response, which combines all accumulated text
into the first chunk and clears all subsequent chunks to "".
The hook then yielded `processed_items[-1]` — the last item, whose
content had been cleared to "". This permanently lost every Nth
chunk's content, causing random missing words/tokens in the client
output (observed in Roo Code, Open WebUI, etc.).
Fix: deep-copy the current chunk before guardrail processing runs,
then yield the original (unmodified) chunk. The guardrail validation
still executes and can block if it detects a problem, but the stream
content is preserved.
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
* fix fail-open for grayswan; pass metadata to cygnal api endpoint; update docs
* pass litellm_metadata to cygnal in payload
* switch error msg to const, and clean exception handling.
* update pyproject.toml as requested
* Revert "update pyproject.toml as requested"
This reverts commit 4eece154d0.
Previously, when Model Armor guardrail blocked a request/response,
the `applied_guardrails` field was not populated in the logs because
`add_guardrail_to_applied_guardrails_header()` was called after the
HTTPException was raised.
This fix moves the `add_guardrail_to_applied_guardrails_header()` call
to before the blocking check in all hooks:
- async_pre_call_hook (pre_call mode)
- async_moderation_hook (during_call mode)
- async_post_call_success_hook (post_call mode)
- async_post_call_streaming_iterator_hook (streaming)
This ensures that even when a guardrail blocks content, the guardrail
name is properly recorded in the logs for observability.
Added regression tests to verify applied_guardrails is populated when
content is blocked.
Co-authored-by: Cursor <cursoragent@cursor.com>
* [Feat] Add model parameter to Generic Guardrail API
Add model information to guardrail requests, allowing guardrails to make
model-specific security decisions.
Changes:
- Add `model` field to GenericGuardrailAPIInputs TypedDict
- Add `model` field to GenericGuardrailAPIRequest Pydantic model
- Update OpenAI and Anthropic handlers to pass model from request/response
- Add unit tests for model parameter handling
* [Feat] Add model parameter to all guardrail_translation handlers
Extend model parameter support to all guardrail handlers for consistent
implementation across all endpoint types:
- OpenAI Responses API (input/output + streaming)
- OpenAI Image Generation (input only)
- OpenAI Text Completion (input/output)
- OpenAI Text-to-Speech (input only)
- OpenAI Audio Transcription (output only)
- Cohere Rerank (input only)
- Pass-through Endpoints (input/output)
- MCP Server (input only)
This addresses the review feedback requesting consistent model parameter
handling across all guardrail_translation/handler.py files.
---------
Co-authored-by: Igal Boxerman <igal@pillar.security>
* Consolidated change
* fix(prompt_security): update message processing to persist sanitized files and filter for API calls
* fix per krrishdholakia suggestion
- Add `should_wrap_with_default_message` parameter to GuardrailRaisedException
- Update Generic Guardrail API to use clean error messages without wrapper
- When should_wrap_with_default_message=False, exception shows the original
blocked_reason directly (e.g., "pii detected") instead of verbose format
- Update test to verify GuardrailRaisedException is raised with clean message
* fix(generic-guardrail-api): fix SerializationIterator error on multimodal requests
When sending multimodal messages (with images) through the Generic Guardrail API,
the `model_dump()` call fails with "Object of type SerializationIterator is not
JSON serializable" error.
Root cause: The `ChatCompletionAssistantMessage` type defines `content` as an
`Iterable` (not just `List`), and Pydantic's `model_dump()` creates a
`SerializationIterator` for iterables which is not JSON serializable.
Fix: Use `model_dump(mode="json")` which properly converts all iterables to
lists and ensures all complex objects are JSON serializable.
* fix(guardrails): pass tools (function definitions) to guardrail inputs
The unified guardrail handler was not passing the `tools` parameter
(function definitions) from the request to the guardrail inputs.
This meant guardrails could not inspect or validate tool definitions.
Added extraction of `data.get("tools")` and inclusion in the
GenericGuardrailAPIInputs passed to `apply_guardrail()`.
* test(guardrails): add tests for tools passed to guardrail
Added tests verifying that tools (function definitions) are correctly
passed to guardrails in the unified guardrail handler:
- test_tools_passed_to_guardrail
- test_multiple_tools_passed_to_guardrail
- test_no_tools_in_request
- test_tools_and_tool_calls_both_passed
* init guardrails
* init guardrails
* some fixes
* some fixes
* ruff
* some fixes
* some fixes
* some fixes
* some fixes
* some fixes
* some fixes
* docs
