AI Security

AI Supply Chain, LLM Vulnerability & Model Security Analysis — 12 Detection Rules

AI Findings

18

Source: AISecurity

Critical

5

Prompt injections + insecure loading

AI Assets Discovered

23

LLMs, agents, pipelines

Compliance

78%

OWASP LLM Top 10 coverage

Detection Rules (12)

Critical Severity

Prompt Injection

Critical

CWE-77 — Detects direct and indirect prompt injection attacks in LLM inputs

Insecure Output Handling

Critical

CWE-116 — Unvalidated LLM output used in downstream systems

Insecure Plugin/Tool Loading

Critical

CWE-829 — Untrusted plugin or tool execution without sandboxing

Model Deserialization Attack

Critical

CWE-502 — Insecure deserialization of ML model files (pickle, ONNX)

High / Medium Severity

Training Data Poisoning

High

CWE-1284 — Tampered or malicious data in training/fine-tuning pipelines

Model Supply Chain Compromise

High

CWE-494 — Unverified model artifacts from external registries

Sensitive Data in Prompts

High

CWE-200 — PII, secrets, or credentials passed to LLM providers

Excessive Agency

High

CWE-269 — LLM agents with overly broad permissions or tool access

RAG Data Leakage

Medium

CWE-212 — Retrieval-augmented generation exposing restricted documents

Embedding Inversion

Medium

CWE-327 — Vector embeddings that can be reversed to recover source text

API Key Exposure in AI Config

High

CWE-798 — Hardcoded API keys for AI providers in configuration files

MCP Server Misconfiguration

Medium

CWE-1188 — Insecure Model Context Protocol server settings or endpoints

Asset Name	Type	Provider	Version	Risk Level	Vulnerabilities
GPT-4o	LLMModel	OpenAI	2024-08-06	Medium	2
Claude 3.5 Sonnet	LLMModel	Anthropic	2024-10-22	Low	None
LangChain	AgentFramework	LangChain Inc.	0.3.14	High	4
FAISS Vector Store	VectorDatabase	Meta	1.8.0	Medium	1
Custom MCP Server	MCPServer	Internal	1.2.0	High	3
Semantic Kernel	AgentFramework	Microsoft	1.44.0	Low	None
Whisper	LLMModel	OpenAI	v3-large	Medium	1
ChromaDB	VectorDatabase	Chroma	0.5.23	Medium	2

Code	Category	Rules Mapped	Status
LLM01	Prompt Injection	2	Covered
LLM02	Insecure Output Handling	1	Covered
LLM03	Training Data Poisoning	1	Covered
LLM04	Model Denial of Service	0	Gap
LLM05	Supply Chain Vulnerabilities	2	Covered
LLM06	Sensitive Information Disclosure	2	Covered
LLM07	Insecure Plugin Design	1	Covered
LLM08	Excessive Agency	1	Covered
LLM09	Overreliance	0	Gap
LLM10	Model Theft	1	Partial

Findings by Category

Risk Distribution by Asset Type

Severity	Title	Category	File	Line	Status
Critical	Direct prompt injection in chat endpoint	Prompt Injection	src/ChatService.cs	142	Open
Critical	Indirect prompt injection via RAG context	Prompt Injection	src/RagPipeline.cs	87	Open
Critical	Pickle deserialization of untrusted model	Insecure Loading	ml/ModelLoader.py	34	In Review
High	OpenAI API key hardcoded in config	API Key Exposure	appsettings.json	18	Open
High	LangChain agent with unrestricted tool access	Excessive Agency	src/AgentConfig.cs	56	In Review
High	PII data sent to external LLM provider	Data Leakage	src/PromptBuilder.cs	203	Open
Medium	RAG pipeline exposing restricted documents	Data Leakage	src/RagService.cs	91	Resolved
Medium	MCP server endpoint without authentication	MCP Misconfiguration	mcp/server.json	12	Open