Threat Database

A structured collection of security threats targeting LLMs, ML pipelines, and AI infrastructure, with an emphasis on practical attack vectors and defensive awareness.

LLM & RAG Security

• Indirect Prompt Injection: How hackers break into RAG through "poisoned" documents.
• Invisible Text Attacks: Hidden instructions in PDFs (white font, zero size) and how to find them.
• The "Ignore Previous Instructions" Vulnerability: Analysis of classical attacks and defense methods.
• Roleplay & Jailbreaking: Attacks like "DAN" (Do Anything Now) and "Developer Mode" — detection in prompts.
• System Prompt Leakage: How to force a model to issue its instructions ("Reveal your system prompt" attacks).
• HTML Comment Injection: Why hidden comments in web pages are dangerous for LLMs.
• Output Constraining Attacks: Attacks that force the model to respond in JSON/XML format to bypass filters.
• Context Window Overflow: Spam attacks and garbage data in RAG pipelines.
• Multilingual Jailbreaks: Why attacks in Russian/Chinese bypass English filters.
• Base64 Obfuscation: Detection of encoded payloads in prompts.
• Steganography in AI (LSB): How malware and hidden payloads are concealed within image pixels (Least Significant Bit).
• Agent Tool Calling Hijack (MCP): How attackers exploit the Model Context Protocol to achieve RCE via autonomous agents.

Model & Supply Chain Security

• Python Pickle RCE: Why downloading .pkl files results in remote code execution.
• PyTorch Malware: The danger of torch.load and hidden instructions in model weights.
• Keras Lambda Injection: Malicious code inside the configuration of neural network layers.
• YAML Deserialization Attacks: Vulnerabilities in configs (yaml.load vs safe_load).
• Typosquatting in Python: Attacks through fake packages (for example, tourch instead of torch).
• Dependency Confusion: How attackers inject their code through requirements.txt.
• Data Poisoning via Malicious URLs: Links to .exe and .sh inside CSV/Parquet datasets.
• Git LFS Pointer Attacks: Substitution of models for pointer text files.
• GGUF & SafeTensors Metadata Exploitation: Bypassing safe serialization formats using XSS and malicious JSON headers.

Secrets & Credential Leaks

• AWS IAM Key Leakage: Risks of AKIA key exposure in Jupyter Notebooks.
• Hugging Face Token Exposure: What are the dangers of Write tokens in public repositories?
• OpenAI API Key Leaks: How to protect your budget from key theft (sk-...).
• SSH Private Key Leaks: Detection of forgotten keys (id_rsa) in datasets.
• Slack & Discord Webhook Security: The risks of phishing through leaked webhooks.
• Google Cloud Credentials: Search for service accounts (.json keys) in the code.
• Generic API Key Detection: How to find non-standard secrets using Shannon entropy analysis.
• Environment Variable Leaks: The danger of .env and .pypirc files in the repository.

Infrastructure & Malicious Activity (For DevOps)

• Zip Bombs & Tar Bombs: Archive decompression attacks leading to DoS and Zip Slip (path traversal) in AI pipelines.
• Cryptojacking in ML Containers: Detection of miners (XMRig, Ethminer) in Docker images.
• Reverse Shell Detection: Search for backdoors in scripts (/bin/sh, nc -e).
• SSRF (Server-Side Request Forgery): Attacks on cloud metadata (169.254.169.254).
• Ransomware Indicators: Signs of cryptographers in Python scripts.
• Dangerous System Calls: Why os.system and subprocess in ML code are a red flag.
• Exfiltration via Curl/Wget: How hackers exfiltrate data from closed learning loops.
• Agentic SSRF & Cloud Metadata Exploitation: How attackers hijack autonomous agent tool calling to exfiltrate AWS/GCP credentials via TOCTOU DNS Rebinding, and how to secure LLM network tools.