NeMo Guardrails — The Cyber Toolchain

NeMo Guardrails v0.24.0

Wed, 26 Aug 2026 19:31:04 GMT

NeMo Guardrails v0.24.0 adds RailOutcome, rail manifests, F5 integration, health endpoints, and expands IORails to 59 built-in rails. • Adds `RailOutcome` — an engine-neutral allow, block, or transform contract returned by rail actions and enforced by both `LLMRails` and `IORails`. • Adds typed rail manifests so built-in rails can declare their configuration, actions, execution surfaces, requirements, and privacy properties for automatic discovery by both engines. • Adds output-rail checking mode to the `/v1/checks` endpoint, which now selects a server-loaded configuration via `config_id` or the server default. • Adds `/v1/health` and `/healthz` server health-check endpoints. • Adds multiple namespaced self-check rails with per-rail namespaced task prompts, enabling more than one self-check rail to run in the same configuration. • Adds F5 Guardrails integration to the built-in rail library. • Adds a canonical outbound HTTP client with connection pooling, retries, TLS, error handling, lifecycle management, and privacy-safe tracing and metrics for all built-in integrations. • Adds shared model telemetry and an instrumented model decorator for LLM call observability. • Expands `IORails` to execute 59 of the library's 67 action-backed input and output surfaces, including community integrations and content-transforming rails, sharing implementations with `LLMRails`. • Adds message checks to `IORails`. • Non-streaming generate() and generate_async() calls on `IORails` now return `GenerationResponse` when generation options are supplied. Breaking changes: • Custom actions using @action(output_mapping=...) must remove that argument and return an explicit `RailOutcome` instead. • The `hf-classifier` install extra is removed; install `transformers` and `torch` directly to use the local classifier backend. • `IORails` calls that supply generation options now return `GenerationResponse`; message lists must be passed with `messages=` as a keyword argument rather than positionally. • Custom Colang 1 flows using the former space-separated Cleanlab, Fiddler, or GCP action names must switch to their `snake_case` names. • `/v1/checks` no longer accepts inline configuration; select a server-loaded configuration via `config_id` or rely on the server default. • Chat Completions requests must use supported role-specific OpenAI message shapes; internal event payloads, unexpected fields, and audio requests are now rejected. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.23.0

Wed, 01 Jul 2026 16:22:19 GMT

NeMo Guardrails v0.23.0 adds tool-call rails, HF classifier rails, PII masking, context bloat detection, and a /v1/checks endpoint. • Adds streaming and non-streaming tool calling to `IORails`, with local rails that validate model-emitted tool calls and application-returned tool results. • New `/v1/checks` endpoint on the OpenAI-compatible server runs input or output rails standalone, returning `passed`, `modified`, or `blocked` status without generating a model response. • Adds lightweight Hugging Face classifier rails for input, output, and retrieval, backed by local Transformers, vLLM, KServe, and FMS. • Adds context bloat detection rails to flag oversized, repetitive, low-entropy, or padded input and retrieved content. • Adds Polygraf integration for PII detection and masking across input, output, and retrieval rails. • Replaces Annoy with exact NumPy search as the default embedding index, removing the native C++ dependency while preserving similarity-threshold semantics. • Extends `IORails` OpenTelemetry support with opt-in content capture and richer request, response, and token-usage span attributes. • LangChain integration adds support for the OpenAI Responses API and Harmony response format models. • Distribution wheel is now approximately ten times smaller by excluding examples and development files. Breaking changes: • Pydantic `>=2.5,<3.0` is now required; environments pinned to Pydantic 1.x must upgrade before installing v0.23.0. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.22.0

Fri, 22 May 2026 20:25:31 GMT

IORails adds streaming, OpenTelemetry, and speculative generation; LangChain becomes optional with NEMOGUARDRAILS_LLM_FRAMEWORK. • `NEMO_GUARDRAILS_NO_USAGE_STATS` environment variable to disable anonymous telemetry collection • `.config/nemoguardrails/do_not_track` file to opt out of usage reporting without setting an environment variable • `NEMOGUARDRAILS_LLM_FRAMEWORK=langchain` environment variable to re-enable LangChain when using non-OpenAI providers • `nemoguardrails.testing` public testing surface for library-level test utilities • Speculative generation in IORails that runs input rails in parallel with the main LLM response to hide latency • IORails streaming output rail support for real-time guardrail evaluation on LLM outputs • IORails OpenTelemetry tracing and token-level metrics integration Breaking changes: • LangChain and LangChain-providers are removed from core dependencies; they must be installed separately when `NEMOGUARDRAILS_LLM_FRAMEWORK=langchain` is set • `AsyncWorkQueue` is moved from the `Guardrails` module to `IORails`, breaking existing imports from the previous location Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.20.0

Thu, 22 Jan 2026 08:44:24 GMT

Adds reasoning guardrail connector and GLiNER for open PII detection • Add support to auto select multilingual refusal bot messages • Adding GLiNER for PII detection (open alternative to PrivateAI) Breaking changes: • The `streaming` field has been removed from the config Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.19.0

Wed, 03 Dec 2025 17:27:00 GMT

Adds LangChain 1.x support with content blocks API and benchmarking scripts • `langchain` v1 compatibility is now supported in NeMo Guardrails. • Adds `llm` provider support for LangChain 1.x content blocks to handle reasoning traces and tool calls. • Adds a `Procfile` to run Guardrails with mock LLMs for benchmarking. • Adds an `AIPerf` run script for benchmarking workflows. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.18.0

Thu, 06 Nov 2025 20:06:41 GMT

New BotThinking events, embedding providers (Azure, Cohere, Google), Cisco AI Defense integration, and in-memory caching for content models. • Introduces `BotThinking` events to apply guardrails on an LLM's reasoning traces • Adds Azure OpenAI as a supported embedding provider • Adds Cohere embedding integration • Adds Google embedding integration • Adds Cisco AI Defense integration • Adds in-memory LFU caching for content-safety, topic-control, and jailbreak detection models to reduce latency • Adds automatic provider inference for LangChain LLMs • Adds custom HTTP header support for the ChatNVIDIA provider Breaking changes: • Reasoning traces are extracted to a separate `reasoning` field instead of being prepended; use LangChain's `additional_kwargs` structure. • `stream_async` now raises an error if output rail streaming is disabled. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.17.0

Thu, 09 Oct 2025 16:45:35 GMT

NeMo Guardrails v0.17.0 adds LangGraph integration, tool rails, and Trend Micro Vision One AI integration • `LLMRails` now supports tool call passthrough for building safe multi-agent workflows with LangGraph. • `RunnableRails` has been rewritten to support the full LangChain Runnable protocol (sync, async, streaming, batch). • New `tool-rails` subsystem provides tool input rails for message validation and processing. • New `tool-rails` subsystem provides tool output rails and validation logic. • Community integration added for Trend Micro Vision One AI Scanner (`AI Guard`) to detect harmful content, data leakage, and prompt injections. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.15.0

Fri, 08 Aug 2025 11:22:08 GMT

NeMo Guardrails v0.15.0 adds parallel rails execution, OpenTelemetry tracing API, and external async token generators. • Update tracing to use otel api • Support parallel rails execution • Implement parallel streaming output rails execution • Support external async token generators Breaking changes: • The tracing subsystem now uses the otel api instead of its previous implementation, requiring migration for existing tracing setups. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.14.1

Tue, 08 Jul 2025 18:50:24 GMT

Adds direct API key configuration support for jailbreak detection • Adds direct API key configuration support for jailbreak detection Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.14.0

Wed, 11 Jun 2025 18:58:04 GMT

New YARA-based output guardrail for code injection detection and Clavata community integration • Adds `fuzzy find provider` capability to the CLI (`nemo-guardrails providers fuzzy-find `) allowing practitioners to locate available guardrail providers via shell command. • Adds inline YARA rules support in the output guardrail configuration (setting `type: yara_inline` and providing `rule: `), enabling direct code-injection detection without external files. • Adds a new community integration for Clavata LLM content moderation (provider type `clavata`) accessible via library configuration or CLI provider selection. • Adds support for preserving and applying guardrails to reasoning traces, controlled by the config key `preserve_reasoning_traces` (boolean) in Rail YAML definitions. Breaking changes: • The config key `remove_thinking_traces` is renamed to `remove_reasoning_traces`; existing YAML files using the old name will fail validation or ignore the setting. • The deprecated argument `return_context` is removed; code passing this argument to library functions will raise an error. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.13.0

Tue, 25 Mar 2025 18:25:34 GMT

NeMo Guardrails v0.13.0 adds multimodal rails, Fiddler integration, and reasoning trace support • `FiddlerGuardrail` integration enables low-latency hosted guardrail models for content safety and hallucination detection. • Multimodal input and output rails allow checking safety of image prompts and vision-language model outputs. • New `NemoGuardJailbreakDetect` NIM support for jailbreak detection capabilities. • Generation metadata is now included in streaming chunks via the `metadata` field on chunk objects. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.12.0

Wed, 26 Feb 2025 18:10:50 GMT

NeMo Guardrails v0.12.0 adds output rails streaming, Prompt Security integration, and score thresholds to AnalyzerEngine • Support Output Rails Streaming • Add pii masking capability to PrivateAI integration • Add embedding_params to BasicEmbeddingsIndex • Add score threshold to AnalyzerEngine Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.11.1

Thu, 16 Jan 2025 22:25:48 GMT

NeMo Guardrails v0.11.1 adds ContentSafety, TopicControl, and JailbreakDetect NIM connectors • `ContentSafety` connector: new integration for the NVIDIA NemoGuard content safety NIM microservice to detect biased or harmful outputs. • `TopicControl` connector: new integration for the NVIDIA NemoGuard topic control NIM microservice to keep conversations on approved topics. • `JailbreakDetect` connector: new integration for the NVIDIA NemoGuard jailbreak detection NIM microservice to identify jailbreak attempts. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.10.0

Fri, 27 Sep 2024 21:47:53 GMT

New Content Safety Module, Colang 2 Migration Tool, LLMRails serialization, and Cleanlab score in NeMo Guardrails v0.10.0 • Adds a `ContentSafetyModule` to integrate content safety models • Introduces a `MigrationTool` for converting Colang 1.0 or 2-alpha configs to Colang 2.x • Enables Rails Exception Handling in Colang 1.0 and 2.x • Adds `relevance_chunks` support to Colang 2 prompts • Implements serialization support for the `LLMRails` class via `llmrails` • Integrates Cleanlab's Trustworthiness Score into the library • Adds GCP moderation support Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.9.1

Thu, 25 Jul 2024 12:02:16 GMT

Adds `--default-config-id` CLI flag, custom embedding model support, and Got It AI / AutoAlign / Patronus Lynx integrations • Adds `--default-config-id` argument to the server CLI for specifying a default config • Enables registering and using custom embedding models via new library APIs • Adds Got It AI truthchecking integration for RAG applications • Adds AutoAlign guardrail integration • Adds Patronus Lynx integration Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.9.0

Fri, 10 May 2024 07:51:27 GMT

Adds Colang 2.0 language and runtime with parallel flows, the ... generation operator, and async actions. • `_generation` operator (`...`) for constructing output sequences within Colang 2.0 flows • Parallel flow execution engine supporting multiple simultaneous concurrent flows • Asynchronous action execution model allowing non-blocking action invocation in Colang 2.0 Breaking changes: • Colang 1.0 syntax and semantics are no longer the default; Colang 2.0 (Beta) is now included, requiring migration for existing configurations. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.8.2

Mon, 01 Apr 2024 20:52:04 GMT

Adds Vertex AI and NVIDIA AI Endpoints integrations; exposes `verbose` arg in `RunnableRails`. • Add support for Vertex AI Models into Guardrails • Add support for NVIDIA AI Endpoints • Expose the `verbose` parameter through `RunnableRails` Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.8.1

Fri, 15 Mar 2024 10:32:20 GMT

Adds streaming example for custom actions in Colang 2.0 • Adds an example demonstrating how to use streaming with a custom action in NeMo Guardrails. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.8.0

Wed, 28 Feb 2024 15:19:09 GMT

NeMo Guardrails v0.8.0 adds jailbreak heuristics, generation options, multi-config API calls, and Colang 2.0 preview • Adds a new input rail type based on `jailbreak heuristics` for detecting adversarial prompts. • Adds `generation_options` to control which rails trigger, what data is returned, and logging details in the response. • Adds support for making API calls to the guardrails server using `multiple configuration ids`. • Adds configurable `stop tokens` per task template. • Adds a preview implementation of `Colang 2.0` for early testing. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.7.0

Wed, 31 Jan 2024 14:54:16 GMT

NeMo Guardrails v0.7.0 adds Llama Guard moderation, server-side threads, and RunnableRails LangChain integration • `RunnableRails` class for improved LangChain integration allowing direct use within LangChain chains. • Support for server-side threads in the NeMo Guardrails library API. • Native support for Llama Guard input and output content moderation as a new guardrail type. Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.6.1

Wed, 20 Dec 2023 22:13:03 GMT

Adds `--version` CLI flag to NeMo Guardrails • Support for `--version` flag in the CLI Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails

NeMo Guardrails v0.6.0

Wed, 13 Dec 2023 21:59:34 GMT

NeMo Guardrails v0.6.0 adds rails configs, AlignScore fact-checking, streaming, and Presidio detection • Adds `rails.dialog.single_call.enabled` config to reduce latency by combining guardrails and response generation in a single LLM call • Integrates ActiveFence ActiveScore API as an input rail for the first time • Adds Presidio integration for sensitive data detection guardrails • Adds LLMRails.explain() method to get explanation of guardrail behavior • Supports AlignScore for fast fact-checking of LLM responses • Adds support for streaming LLM responses when no output rails are configured • Adds `prompting_mode` config key to control prompting modes • Enables loading configuration from a dictionary using RailsConfig.from_content(config=...) • Adds Dockerfile for building a container image Full history on Cyber Toolchain: https://cybertoolchain.io/tools/nemo-guardrails