How It Works

LLM Armor applies two complementary analysis layers to every Python file.

Layer 1: Regex

Fast line-by-line pattern matching for common vulnerability patterns. Runs on all files regardless of whether they parse as valid Python.

Layer 2: AST taint analysis

Python’s ast module parses each file into a syntax tree and performs source-based taint tracking. This catches patterns regex cannot detect: variable aliasing, role-aware dict construction, multi-line string concatenation, and **kwargs dict spreading.

If a file has syntax errors, the AST layer falls back gracefully, leaving regex results intact.

Deduplication

When both layers detect the same issue on the same line, only one finding is reported.

Taint sources

Tainted (user-controlled)	Example
HTTP request	data = request.json["prompt"]
HTTP form	data = request.form.get("field")
Django request	data = request.POST["query"]
stdin	data = input("Enter: ")
CLI arguments	data = sys.argv[1]
WebSocket	data = websocket.receive()
Function parameter	def handle(user_msg):
@tool parameter	@tool def my_tool(command: str):

Safe sources (not tainted)

Source	Example
Config lookup	prompt = config.get("default_prompt")
Environment variable	prompt = os.environ["PROMPT"]
Database call	prompt = db.fetch_prompt(id)
String literal	prompt = "You are a helpful assistant."

Taint propagates through direct alias assignments but not through function calls, so clean = sanitize(raw) does not taint clean.