Feature/openai llm

.gitignore

@@ -160,3 +160,5 @@ cython_debug/
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 .idea/
 
+# Custom data folder for vectordb and sigma rules
+sigmaiq/llm/data/*

README.md

@@ -33,6 +33,9 @@ encountered. Please report any issues [here](https://github.com/AttackIQ/SigmAIQ
 Feature requests are also always welcome! pySigma tools/utils are currently not in the pre-release version,
 and will be added in future releases.
 
+# LLM Support
+For LLM usage, see the [LLM README](sigmaiq/llm/README.md)
+
 # Installation & Usage
 
 ## Installation

examples/custom_field_mappings.py

@@ -0,0 +1,49 @@
+# %% This example shows how to use the SigmAIQ pySigma wrapper to provide custom field mappings for a backend
+# %% This will allow you to translate specific field names to custom field names during rule translation
+
+# %% Import SigmAIQ
+from sigmaiq import SigmAIQBackend, SigmAIQPipeline
+
+# %% Import pprint for pretty printing, and copy for copying rules
+from pprint import pprint
+from copy import copy
+
+# %% A basic Sigma Rule in YAML str to convert to a query.
+# %% SigmAIQ also accepts a rule in JSON/Dict format, SigmaRule objects, and SigmaCollection objects
+
+sigma_rule = """
+title: whoami Command
+description: Detects a basic whoami commandline execution
+logsource:
+    product: windows
+    category: process_creation
+detection:
+    selection1:
+        - CommandLine|contains: 'whoami.exe'
+    condition: selection1
+"""
+
+# %% Create SigmAIQ backend translate the rule to a Microsoft 365 Defender query
+sigmaiq_backend = SigmAIQBackend(backend="splunk").create_backend()
+query = sigmaiq_backend.translate(copy(sigma_rule))  # Returns List of queries
+
+print("\nM365Defender Query: ", end="\n\n")
+pprint(query[0])
+print("\n-------------------")
+
+# %% Create custom field mappings
+# %% This will map the CommandLine field to a custom field named "CustomCommandLine"
+custom_field_mappings = {"CommandLine": "CustomCommandLine"}
+my_custom_pipeline = SigmAIQPipeline.from_fieldmap(custom_field_mappings, priority=0).create_pipeline()
+
+# %% Create SigmAIQ backend translate the rule to a Microsoft 365 Defender query with our custom field mappings
+sigmaiq_backend = SigmAIQBackend(
+    backend="splunk",
+    processing_pipeline=my_custom_pipeline).create_backend()
+
+query = sigmaiq_backend.translate(copy(sigma_rule))  # Returns List of queries
+
+print("\nM365Defender Query with Custom Fieldmappings: ", end="\n\n")
+pprint(query[0])
+print("\n-------------------")
+

examples/llm_basic_usage.py

@@ -0,0 +1,36 @@
+# %% This example will demonstrate how to use SigmAIQ to perform the following:
+# %% 1. Download the latest Sigma Rule package release
+# %% 2. Create embeddings of the Sigma Rules in the package
+# %% 3. Create and save a VectorDB of the Sigma Rule embeddings
+# %% 4. Use a similarity search on the VectorDB to find Sigma Rules similar to a provided query
+from pprint import pprint
+
+# %% NOTE, this example uses OpenAI for embeddings. Ensure you have an OpenAI API key set in your environment variable
+# %% OPENAI_API_KEY
+
+# %% Also ensure you have installed the correct requirements with:
+# `pip install -r requirements/common.txt -r requirements/llm.txt`
+
+
+# %% Import SigmAIQ LLM and OpenAIEmbeddings
+from sigmaiq.llm.base import SigmaLLM
+
+# %% Create a SigmaLLM object with default settings. See the class docstring for more information
+sigma_llm = SigmaLLM()
+
+# %% The `create_sigma_vectordb()` method will automatically do all the work for you :) (only run this once)
+sigma_llm.create_sigma_vectordb(save=True)  # Save locally to disk
+
+# %% Run a similarity search on the vectordb for encoded powershell commands and print top 3 results
+query = "Encoded powershell commands"
+matching_rules = sigma_llm.simple_search(query, k=3)
+for matching_rule in matching_rules:
+    print(matching_rule.page_content, end="\n\n-------------------\n\n")
+
+# %% You can also load an existing vector store from disk (recommended)
+sigma_llm.load_sigma_vectordb()
+
+query = "certutil"
+matching_rules = sigma_llm.simple_search(query, k=3)
+for matching_rule in matching_rules:
+    print(matching_rule.page_content, end="\n\n-------------------\n\n")

examples/llm_rule_translation_and_creation.py

@@ -0,0 +1,58 @@
+# %% This example will demonstrate how to create a Sigma langchain agent chatbot, which can perform various tasks like
+# %% automatically translate a rule for you, and create new rules from a users input.
+
+# %% Import required SigmAIQ classes and functions
+from sigmaiq.llm.toolkits.base import create_sigma_agent
+from sigmaiq.llm.base import SigmaLLM
+
+# %% Ensure we have our Sigma vector store setup with our base LLM class
+sigma_llm = SigmaLLM()
+
+try:
+    sigma_llm.load_sigma_vectordb()
+except Exception as e:
+    print(e)
+    print("Creating new Sigma VectorDB")
+    sigma_llm.create_sigma_vectordb(save=True)
+
+# %% Create a Sigma Agent Executor, and pass it our Sigma VectorDB
+sigma_agent_executor = create_sigma_agent(sigma_vectorstore=sigma_llm.sigmadb)
+
+# %% RULE TRANSLATION
+# %% Have the agent automatically translate a Sigma rule to a Splunk query with the splunk_cim_dm pipeline
+
+sigma_rule = r"""
+title: whoami Command
+description: Detects a basic whoami commandline execution
+logsource:
+    product: windows
+    category: process_creation
+detection:
+    selection1:
+        - CommandLine|contains: 'whoami.exe'
+    condition: selection1
+"""
+
+user_input = ("Translate the following Sigma rule to a Splunk query using the 'splunk_cim_dm' pipeline: \n\n" +
+              sigma_rule)
+
+# answer = sigma_agent_executor.invoke({"input": user_input})
+# print("\nRULE TRANSLATION:", end="\n\n")
+#print(f"Question:\n {user_input}", end="\n\n")
+#print(f"Answer: \n")
+#print(answer.get('output'), end="\n\n")
+
+# %% RULE CREATION
+# %% The agent will take the user input, look up similar Sigma Rules in the Sigma vector store, then create a brand
+# %% new rule based on the context of the users input and the similar Sigma Rules.
+
+user_input = ("Create a Windows process creation Sigma Rule for certutil downloading a file "
+              "from definitely-not-malware.com, then translate it to a Microsoft 365 Defender query.")
+
+answer = sigma_agent_executor.invoke({"input": user_input})
+print("\nRULE CREATION:", end="\n\n")
+print(f"Question:\n {user_input}", end="\n\n")
+print(f"Answer: \n")
+print(answer.get('output'), end="\n\n")
+
+

examples/translate_sigma_rules.py

@@ -0,0 +1,77 @@
+# %% This example shows how to use the SigmAIQ pySigma wrapper to easily translate Sigma rules to queries
+# %% easily, without having to worry about installing and configuring the correct backends, pipelines and other details.
+
+
+# %% Import SigmAIQ
+from sigmaiq import SigmAIQBackend
+
+# %% Import pprint for pretty printing, and copy for copying rules
+from pprint import pprint
+from copy import copy
+
+# %% A basic Sigma Rule in YAML str to convert to a query.
+# %% SigmAIQ also accepts a rule in JSON/Dict format, SigmaRule objects, and SigmaCollection objects
+
+sigma_rule = """
+title: whoami Command
+description: Detects a basic whoami commandline execution
+logsource:
+    product: windows
+    category: process_creation
+detection:
+    selection1:
+        - CommandLine|contains: 'whoami.exe'
+    condition: selection1
+"""
+
+# %% BACKENDS
+# %% Show the available supported backends
+print("Supported Backends:", end="\n\n")
+pprint(SigmAIQBackend.display_available_backends())
+print("\n-------------------")
+
+# %% Create SigmAIQ backend translate the rule to a Microsoft 365 Defender query
+# %% SigmAIQ will automatically select the best pipeline for the backend
+sigmaiq_backend = SigmAIQBackend(backend="microsoft365defender").create_backend()
+query = sigmaiq_backend.translate(copy(sigma_rule))  # Returns List of queries
+
+print("\nM365Defender Query: ", end="\n\n")
+pprint(query[0])
+print("\n-------------------")
+
+# %% PIPELINES
+# %% Show the available pipelines with each backend
+print("Available Pipelines:", end="\n\n")
+pprint(SigmAIQBackend.display_all_associated_pipelines())
+print("\n-------------------")
+
+# %% Create SigmAIQ backend translate the rule to a Splunk search with the CIM pipeline
+sigmaiq_backend = SigmAIQBackend(backend="splunk", processing_pipeline="splunk_cim_dm").create_backend()
+query = sigmaiq_backend.translate(copy(sigma_rule))
+
+print("\nSplunk CIM Query: ", end="\n\n")
+pprint(query[0])
+print("\n-------------------")
+
+# %% OUTPUT FORMATS
+# %% Show the available output formats with each backend
+print("\nAvailable Output Formats:", end="\n\n")
+pprint(SigmAIQBackend.display_backends_and_outputs())
+print("\n-------------------")
+
+# %% Change the output_format to an Enterprise Security Correlation Search stanza
+sigmaiq_backend.set_output_format("stanza")
+query = sigmaiq_backend.translate(copy(sigma_rule))
+
+print("\nSplunk CIM Query, Stanza Output: ", end="\n\n")
+pprint(query[0])
+print("\n-------------------")
+
+
+# %% You can also translate a Sigma rule to all supported backend, pipeline, and output format combinations at once.
+# %% Any combination that is not supported will not be included in the results
+# %% This is useful for testing and comparing the output of different backends and pipelines
+queries = SigmAIQBackend.create_all_and_translate(copy(sigma_rule))
+
+print("\n All Translations: ", end="\n\n")
+pprint(queries)

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "SigmAIQ"
-version = "0.2.4"
+version = "0.3.0"
 description = "Wrapper and tools for pySigma and Sigma rules"
 authors = ["Stephen Lincoln <stephen.lincoln@attackiq.com>", "AttackIQ <support@attackiq.com>"]
 readme = "README.md"
@@ -20,7 +20,7 @@ packages = [
 ]
 
 [tool.poetry.dependencies]
-python = "^3.8"
+python = ">=3.8.1, <=3.11.6"
 pysigma = "0.9.11"
 certifi = "^2023.07.22"
 pysigma-backend-carbonblack = "0.1.4"
@@ -39,16 +39,20 @@ pysigma-pipeline-sysmon = "1.0.2"
 pysigma-pipeline-windows = "1.1.0"
 importlib-resources = "^5.13.0"
 
-
-
-
 [tool.poetry.dev-dependencies]
 pytest = "^7.4.0"
 pytest-cov = "^4.1.0"
 black = "^23.7.0"
 ruff = "^0.0.286"
 
+[tool.poetry.group.llm]
+optional = true
 
+[tool.poetry.group.llm.dependencies]
+langchain = "^0.0.335"
+openai = "^1.2.4"
+tiktoken = "^0.5.1"
+faiss-cpu = "^1.7.4"
 
 [build-system]
 requires = ["poetry-core"]

requirements.txt

@@ -1,25 +1 @@
-certifi==2023.7.22 ; python_version >= "3.8" and python_version < "4.0"
-charset-normalizer==3.2.0 ; python_version >= "3.8" and python_version < "4.0"
-idna==3.4 ; python_version >= "3.8" and python_version < "4.0"
-importlib-resources==5.13.0 ; python_version >= "3.8" and python_version < "4.0"
-packaging==22.0 ; python_version >= "3.8" and python_version < "4.0"
-pyparsing==3.1.1 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-carbonblack==0.1.4 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-cortexxdr==0.1.1 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-elasticsearch==1.0.5 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-insightidr==0.2.1 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-loki==0.9.1 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-microsoft365defender==0.2.1 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-opensearch==1.0.0 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-qradar-aql==0.2.3 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-sentinelone==0.1.2 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-splunk==1.0.2 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-backend-stix==0.1.8 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-pipeline-crowdstrike==1.0.0 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-pipeline-sysmon==1.0.2 ; python_version >= "3.8" and python_version < "4.0"
-pysigma-pipeline-windows==1.1.0 ; python_version >= "3.8" and python_version < "4.0"
-pysigma==0.9.11 ; python_version >= "3.8" and python_version < "4.0"
-pyyaml==6.0.1 ; python_version >= "3.8" and python_version < "4.0"
-requests==2.31.0 ; python_version >= "3.8" and python_version < "4.0"
-urllib3==2.0.4 ; python_version >= "3.8" and python_version < "4.0"
-zipp==3.16.2 ; python_version >= "3.8" and python_version < "3.10"
+-r requirements/common.txt

requirements/common.txt

@@ -0,0 +1,25 @@
+certifi==2023.7.22 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+charset-normalizer==3.3.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+idna==3.4 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+importlib-resources==5.13.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+packaging==22.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pyparsing==3.1.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-carbonblack==0.1.4 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-cortexxdr==0.1.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-elasticsearch==1.0.5 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-insightidr==0.2.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-loki==0.9.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-microsoft365defender==0.2.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-opensearch==1.0.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-qradar-aql==0.2.3 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-sentinelone==0.1.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-splunk==1.0.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-backend-stix==0.1.8 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-pipeline-crowdstrike==1.0.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-pipeline-sysmon==1.0.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma-pipeline-windows==1.1.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pysigma==0.9.11 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pyyaml==6.0.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+requests==2.31.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+urllib3==2.1.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+zipp==3.17.0 ; python_full_version >= "3.8.1" and python_version < "3.10"

requirements/llm.txt

@@ -0,0 +1,60 @@
+aiohttp==3.8.6 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+aiosignal==1.3.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+annotated-types==0.6.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+anyio==3.7.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+async-timeout==4.0.3 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+attrs==23.1.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+backoff==2.2.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+beautifulsoup4==4.12.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+certifi==2023.7.22 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+chardet==5.2.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+charset-normalizer==3.3.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+click==8.1.7 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+colorama==0.4.6 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6" and platform_system == "Windows"
+dataclasses-json==0.6.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+distro==1.8.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+emoji==2.8.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+exceptiongroup==1.1.3 ; python_full_version >= "3.8.1" and python_version < "3.11"
+faiss-cpu==1.7.4 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+filetype==1.2.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+frozenlist==1.4.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+greenlet==3.0.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6" and (platform_machine == "win32" or platform_machine == "WIN32" or platform_machine == "AMD64" or platform_machine == "amd64" or platform_machine == "x86_64" or platform_machine == "ppc64le" or platform_machine == "aarch64")
+h11==0.14.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+httpcore==1.0.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+httpx==0.25.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+idna==3.4 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+joblib==1.3.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+jsonpatch==1.33 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+jsonpointer==2.4 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+langchain==0.0.335 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+langdetect==1.0.9 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+langsmith==0.0.64 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+lxml==4.9.3 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+marshmallow==3.20.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+multidict==6.0.4 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+mypy-extensions==1.0.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+nltk==3.8.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+numpy==1.24.4 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+openai==1.2.4 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+packaging==22.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pydantic-core==2.14.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pydantic==2.5.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+python-iso639==2023.6.15 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+python-magic==0.4.27 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+pyyaml==6.0.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+rapidfuzz==3.5.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+regex==2023.10.3 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+requests==2.31.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+six==1.16.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+sniffio==1.3.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+soupsieve==2.5 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+sqlalchemy==2.0.23 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+tabulate==0.9.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+tenacity==8.2.3 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+tiktoken==0.5.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+tqdm==4.66.1 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+typing-extensions==4.8.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+typing-inspect==0.9.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+unstructured==0.10.30 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+urllib3==2.1.0 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"
+yarl==1.9.2 ; python_full_version >= "3.8.1" and python_full_version <= "3.11.6"