llm-red-team
Fail
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: CRITICALPROMPT_INJECTIONOBFUSCATIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill contains a vast repository of adversarial payloads (e.g., DAN, STAN, AIM variants) and instructions designed to override system prompts and bypass safety filters. These were flagged by automated scanners as potential attempts to gain unrestricted access to LLM systems, though they are presented here as test cases for security assessment.
- [DATA_EXFILTRATION]: Provides example payloads designed to demonstrate data exfiltration, including the capture of session cookies and transmission to external example domains like
attacker.example.com. - [OBFUSCATION]: Documents several techniques used to evade security filters, including Base64 encoding of malicious instructions, homoglyph substitution using confusable Cyrillic characters, and the injection of zero-width Unicode characters between tokens.
- [INDIRECT_PROMPT_INJECTION]: Includes detailed guides and payloads for performing indirect prompt injection attacks, specifically focusing on RAG (Retrieval-Augmented Generation) poisoning and the manipulation of agentic tool outputs to hijack conversation flow.
- [REMOTE_CODE_EXECUTION]: The methodology involves the installation and execution of external security tools and custom Python scripts. While targeting legitimate security research, these patterns represent the execution of remote code at runtime for scanning and exploitation testing.
- [COMMAND_EXECUTION]: Utilizes a wide range of shell commands for infrastructure reconnaissance, model fingerprinting, and running automated security sweeps using tools like
garakandffuf. - [EXTERNAL_DOWNLOADS]: Instructions direct the agent to fetch external code and dependencies from various sources, including official GitHub repositories for security tools like
PurpleLlama,PyRIT, andpromptfoo.
Recommendations
- Contains 1 malicious URL(s) - DO NOT USE
Audit Metadata