LLM Detection Rules

Detection logic for monitoring and identifying abuse patterns in locally hosted LLM deployments.

Covers prompt injection attempts, system prompt leakage, request volume abuse, and unauthorized tool calls — mapped to OWASP Top 10 for LLM Applications 2025.

What's in here

File	What it detects
logger.py	Logs every prompt + response to structured JSONL
detect.py	Runs detection rules against the log file
rules/LLM01-prompt-injection.py	Known injection phrasing in user input
rules/LLM02-volume-abuse.py	Unusual request volume from single user
rules/LLM07-prompt-leakage.py	Responses that resemble system prompt leakage
sample-logs/example.jsonl	Sample log file to test detection against

OWASP Coverage

ID	Risk	Detected
LLM01	Prompt Injection	✅
LLM02	Sensitive Information Disclosure	✅
LLM06	Excessive Agency (unauthorized tool calls)	✅
LLM07	System Prompt Leakage	✅

How to run

# 1. Install dependencies
pip install ollama

# 2. Start logging (wraps your local Ollama model)
python logger.py

# 3. Run detections against the log
python detect.py --log sample-logs/example.jsonl

False positive note

Pattern matching on injection phrases will fire on users legitimately discussing AI security. Known FP sources are documented in each rule file. A detection that volunteers its own false positive rate is more useful than one that does not.

Personal project — independent research