LLM Guard

No entries match these filters. Clear filters

1. 0.3.14 Jun 17, 2024 · release notes

   LLM Guard 0.3.14 adds per-scan scanner suppression and custom uvicorn config; drops gunicorn and caching

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch 0.3.14 https://github.com/protectai/llm-guard.git
   # already have the repo? check out this version:
   $ git checkout 0.3.14

   • Adds the ability to suppress specific scanners during analysis in the API.
   • Allows custom uvicorn configuration in the API deployment.

   2 details in Everything

2. 0.3.12 Apr 23, 2024 · release notes

   LLM Guard 0.3.12 adds BanCode scanner, NoRefusalLight, lazy loading, and POST /scan/prompt API endpoint

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch 0.3.12 https://github.com/protectai/llm-guard.git
   # already have the repo? check out this version:
   $ git checkout 0.3.12

   └──▷ USE IT
   Detect and block code snippets in user prompts using the new scanner
   code copy

   from llm_guard.scanners import BanCode
   scanner = BanCode()
   sanitized_prompt, scores, result = scanner.scan(prompt)

   Scan a prompt without sanitizing it for faster inference via the new API endpoint
   shell copy

   curl -X POST http://localhost:8000/scan/prompt -d '{"prompt": "user input"}'

   Configure Gunicorn to run the LLM Guard API with lazy loading support
   shell copy

   gunicorn -w 4 app:app

   • Adds BanCode scanner to detect and block code snippets in the prompt
   • Adds NoRefusalLight scanner to detect refusal using common phrases
   • Adds lazy loading of models in the API deployment
   • Adds support for gunicorn in the API deployment
   • Adds POST /scan/prompt endpoint to scan prompts without sanitizing them

   +1 more show less

   • Adds local folder model loading support to the API

   6 details in Everything

   └──▷ BREAKING ON UPGRADE

   • !The model_kwargs and pipeline_kwargs parameters are removed as they are now part of the new Model object
3. 0.3.10 Mar 14, 2024 · release notes

   LLM Guard 0.3.10 adds Gibberish scanner, AI4Privacy NER models, and local model loading.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch 0.3.10 https://github.com/protectai/llm-guard.git
   # already have the repo? check out this version:
   $ git checkout 0.3.10

   └──▷ USE IT
   Scan input text to detect if it is gibberish using the new scanner.
   code copy

   from llm_guard.input_scanners import Gibberish
   scanner = Gibberish()
    sanitized_prompt, result, score = scanner.scan(prompt)

   Configure the Anonymize scanner patterns using Python configuration.
   code copy

   from llm_guard.input_scanners import Anonymize
   scanner = Anonymize(patterns=[{"name": "credit_card", "pattern": r'\d{4}-\d{4}-\d{4}-\d{4}'}])
    sanitized_prompt, result, score = scanner.scan(prompt)

   Load a custom or downloaded model from a local directory.
   code copy

   from llm_guard.input_scanners import Anonymize
   scanner = Anonymize(model_path="/local/path/to/model")
    sanitized_prompt, result, score = scanner.scan(prompt)

   • Adds Gibberish scanner to detect gibberish in input text.
   • Adds support for Isotonic/distilbert_finetuned_ai4privacy_v2 and Isotonic/deberta-v3-base_finetuned_ai4privacy_v2 NER models for the Anonymize capability.
   • Enables loading models from local folders instead of requiring HuggingFace access.
   • Allows passing pipeline and model kwargs to scanners for better control.

   4 details in Everything

   └──▷ BREAKING ON UPGRADE

   • !Anonymize scanner patterns are now configured in Python instead of a JSON file.
4. 0.3.9 Feb 8, 2024 · release notes

   LLM Guard 0.3.9 adds API auth, rate limiting, ANONYMIZE zh support, and NoRefusal match_type control.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch 0.3.9 https://github.com/protectai/llm-guard.git
   # already have the repo? check out this version:
   $ git checkout 0.3.9

   └──▷ USE IT
   Scan prompts allowing only sentence-level refusal matches for finer granularity.
   code copy

   NoRefusal(match_type="sentence")

   Enable API authentication with basic auth and rate limiting in config file.
   yaml copy

   auth:
     type: basic
     username: admin
     password: secret
   rate_limit:
     enabled: true
     max_requests: 100

   Anonymize Chinese text to replace PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. entities.
   code copy

   result = Anonymize(text, languages=["zh"])

   • API: Adds HTTP basic authentication and API key authentication endpoints
   • API: Adds rate limiting configuration via auth section in config
   • API: Adds OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. support for tracing and metrics
   • Anonymize: Adds Chinese (zh) language support to Anonymize function
   • Anonymize: Adds new regex patterns PO_BOX_RE, PRICE_RE, HEX_COLOR, TIME_RE, DATE_RE, URL_RE, PHONE_NUMBER_WITH_EXT, BTC_ADDRESS for data masking

   +1 more show less

   • NoRefusal: Adds match_type parameter (values: sentence, all) to control matching scope

   6 details in Everything

   └──▷ BREAKING ON UPGRADE

   • !Code: updated model requires updating the languages parameter to match [philomath-1209/programming-language-identification](https://huggingface.org/philomath-1209/programming-language-identification) vocab
   • !API configuration changes with separate sections for auth, rate_limit and cache
5. 0.3.4 Dec 21, 2023 · release notes

   LLM Guard 0.3.4 adds matching strategies to Deanonymize scanner and GPUGPUA graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. ONNXONNXAn open file format for trained models, so a model built in one framework can run in another runtime or on other hardware. ONNX Runtime is the common engine that executes it. support

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch 0.3.4 https://github.com/protectai/llm-guard.git
   # already have the repo? check out this version:
   $ git checkout 0.3.4

   └──▷ TRY IT
   Run the Deanonymize scanner using the ONNX runtime on a GPU device.
   copy

   $ llm-guard scan --tool deanonymize --device cuda

   • Deanonymize scanner now supports configurable matching strategies.
   • Adds ONNX runtime support on GPU for the Deanonymize scanner, enabling faster inference.

   2 details in Everything

   └──▷ BREAKING ON UPGRADE

   • !The Anonymize scanner no longer uses the dbmdz/bert-large-cased-finetuned-conll03-english model.
6. 0.3.2 Nov 15, 2023 · release notes

   LLM Guard 0.3.2 adds ONNXONNXAn open file format for trained models, so a model built in one framework can run in another runtime or on other hardware. ONNX Runtime is the common engine that executes it. inference support for multiple scanners, improves MaliciousURLs model, and updates Langchain integration.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch 0.3.2 https://github.com/protectai/llm-guard.git
   # already have the repo? check out this version:
   $ git checkout 0.3.2

   • Adds ONNX inference support to BanTopics, NoRefusal, FactualConsistency, and Relevance scanners
   • Switches the MaliciousURLs scanner to use the DunnBC22/codebert-base-Malicious_URLs model
   • Updates the Language and LanguageSame scanners to use the papluca/xlm-roberta-base-language-detection model
   • Provides an up-to-date Langchain integration using LCEL

   4 details in Everything

7. 0.3.1 Nov 9, 2023 · release notes

   Adds use_onnx parameter to enable ONNXONNXAn open file format for trained models, so a model built in one framework can run in another runtime or on other hardware. ONNX Runtime is the common engine that executes it. runtime for faster CPUCPUThe main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware. inference in scanners.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch 0.3.1 https://github.com/protectai/llm-guard.git
   # already have the repo? check out this version:
   $ git checkout 0.3.1

   └──▷ USE IT
   Enable ONNX runtime for faster CPU inference in a specific scanner.
   python copy

   scanner = PromptInjection(use_onnx=True)

   • Enables ONNX runtime support in scanners by adding the use_onnx parameter.
   • Switches PromptInjection scanner to use deepset/deberta-v3-base-injection model.

   2 details in Everything

   └──▷ BREAKING ON UPGRADE

   • !Removes the PromptInjectionV2 scanner; users must now use the single PromptInjection scanner.
8. 0.3.0 Oct 14, 2023 · release notes

   LLM Guard 0.3.0 adds Regex, Language, JSON, and LanguageSame scanners plus configurable embedding modelsembedding modelsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG..

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch 0.3.0 https://github.com/protectai/llm-guard.git
   # already have the repo? check out this version:
   $ git checkout 0.3.0

   • Adds Regex scanner to the prompt
   • Adds Language scanners for both prompt and output
   • Adds JSON output scanner capable of repairing broken JSON
   • Adds LanguageSame output scanner to verify prompt and output languages match
   • Enables configurable embeddingembeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. models in the Relevance scanner

   5 details in Everything

   └──▷ BREAKING ON UPGRADE

   • !The Refutation scanner is renamed to FactualConsistency
   • !Removed chunking behavior in Anonymize and Sensitive scanners which previously affected redaction output
9. 0.2.4 Oct 7, 2023 · release notes

   LLM Guard v0.2.4 adds prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. scanner v2, sub-string redaction via BanSubstrings, and new secrets detectors.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch 0.2.4 https://github.com/protectai/llm-guard.git
   # already have the repo? check out this version:
   $ git checkout 0.2.4

   • Adds BanSubstrings capability to redact found substrings in evaluated content
   • Introduces new secrets detectors for identifying sensitive information in prompts or outputs
   • Adds Langchain LCEL integration example demonstrating expression language usage

   3 details in Everything

10. 0.2.3 Sep 23, 2023 · release notes

    LLM Guard 0.2.3 adds fail_fast flag to stop execution on first failure and updates API/Playground support.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch 0.2.3 https://github.com/protectai/llm-guard.git
    # already have the repo? check out this version:
    $ git checkout 0.2.3

11. 0.2.1 Sep 21, 2023 · release notes

    LLM Guard 0.2.1 adds Regex output scanner with redaction and threshold param to Code scanner.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch 0.2.1 https://github.com/protectai/llm-guard.git
    # already have the repo? check out this version:
    $ git checkout 0.2.1

    └──▷ USE IT
    Redact credit card numbers (16 digits) from LLM output using the new Regex scanner capabilities.
    code copy

    from llm_guard.output_scanners.regex import Regex
    scorer = Regex(r'\d{4}-\d{4}-\d{4}-\d{4}')

    Tune the false-positive rate for code snippet detection by setting a custom threshold.
    code copy

    from llm_guard.input_scanners.code import Code
    scorer = Code(threshold=0.85)

    • Regex output scanner can now redact text using a user-provided regular expression pattern.
    • Adds threshold parameter to Code scanners to control the similarity threshold for detection.

    2 details in Everything

12. 0.2.0 Sep 15, 2023 · release notes

    LLM Guard 0.2.0 adds GPUGPUA graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. inference support and Streamlit demo app

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch 0.2.0 https://github.com/protectai/llm-guard.git
    # already have the repo? check out this version:
    $ git checkout 0.2.0

    • Adds GPU inference support for scanners to accelerate processing
    • Includes a Streamlit demo application deployed to HuggingFace Spaces

    2 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The URL entity type is removed from the Anonymize scanner, which may cause false-negative results for URLs that were previously detected.
13. 0.1.0 Aug 12, 2023 · release notes

    LLM Guard v0.1.0 introduces Refutation, MaliciousURLs, Secrets scanners with risk scoring and faker support

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch 0.1.0 https://github.com/protectai/llm-guard.git
    # already have the repo? check out this version:
    $ git checkout 0.1.0

    • Adds Refutation output scanner for factual consistency checking
    • Adds MaliciousURLs output scanner to detect malicious URLs in model outputs
    • Adds Secrets prompt scanner to detect secrets in prompts
    • Introduces risk score (0-1) for all prompt and output scanners
    • Adds use_faker parameter to Anonymize prompt scanner to use faker instead of placeholders

    +1 more show less

    • Updates Anonymize prompt scanner to use transformer based Spacy model en_core_web_trf

    6 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Anonymize prompt scanner no longer uses FILE_EXTENSION entity type
14. 0.0.1 Aug 7, 2023 · release notes

    LLM Guard 0.0.1 introduces the first input and output scanner classes for detecting prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools., toxicity, and sensitive data in LLM interactions.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch 0.0.1 https://github.com/protectai/llm-guard.git
    # already have the repo? check out this version:
    $ git checkout 0.0.1

    • InputScanners with tests: Anonymize, BanSubstrings, BanTopics, Code, PromptInjection, Sentiment, TokenLimit, and Toxicity
    • OutputScanners with tests: BanSubstrings, BanTopics, Code, Deanonymize, NoRefusal, Regex, Relevance, Sensitive, and Toxicity

    2 details in Everything

15. 0.3.7 Jan 15, 2023 · release notes

    LLM Guard 0.3.7 adds URLReachability, BanCompetitors, InvisibleText, and ReadingTime scanners.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch 0.3.7 https://github.com/protectai/llm-guard.git
    # already have the repo? check out this version:
    $ git checkout 0.3.7

    └──▷ USE IT
    Detect steganography attacks by checking for invisible unicode characters in the input prompt.
    code copy

    scanner = InvisibleText()
    result = scanner.scan_input(prompt)

    • URLReachability scanner to check if the URL is reachable
    • BanCompetitors scanner to check if the prompt or output contains competitors' names
    • InvisibleText scanner to check if the prompt contains invisible unicode characters (steganography attack)
    • ReadingTime scanner to check if the output can be read in less than a certain amount of time

    4 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Regex scanners have a new signature. It accepts patterns, is_blocked and match_type.
    • !BanSubstrings: match_type parameter became Enum instead of str.
    • !Code scanners have a new signature. It accepts languages and is_blocked instead of 2 separate lists.
    • !Using keyword arguments for better readability of the code e.g. scanner = BanSubstrings(["a", "b", "c"], "str", False, True, False) would raise an error.