The Agent Skills Directory

[PROMPT_INJECTION]: The reference material in references/llm-attacks.md provides a wide array of prompt injection payloads and jailbreak templates. These include techniques like the 'DAN' (Do Anything Now) persona, instruction overrides ('Ignore all previous instructions'), and system prompt extraction queries designed to bypass LLM safety filters.
[OBFUSCATION]: The skill implements and describes several obfuscation methods for evading security filters. In references/llm-attacks.md, logic is provided for hiding instructions using zero-width Unicode characters (U+200B, U+200C, U+200D). Additionally, both references/adversarial-ml.md and references/llm-attacks.md demonstrate homoglyph substitution attacks using Cyrillic and Greek characters to bypass text-based detection.
[DYNAMIC_EXECUTION]: In references/model-file-forensics-and-deserialization.md, the skill provides code for creating a malicious Python Pickle artifact. This artifact uses the __reduce__ method to execute arbitrary shell commands via os.system when the file is deserialized, illustrating a remote code execution (RCE) vector.
[COMMAND_EXECUTION]: The skill describes methods for exploiting tool-use capabilities through argument injection. Examples include payloads designed to read sensitive system files (e.g., /etc/passwd, /flag.txt) or execute arbitrary commands by manipulating parameters passed to agent tools.
[INDIRECT_PROMPT_INJECTION]: The skill identifies and provides payloads for indirect prompt injection via poisoned RAG (Retrieval-Augmented Generation) inputs and web pages in references/llm-attacks.md. These inputs lack boundary markers or sanitization, allowing malicious instructions embedded in documents to influence agent capabilities like file reading and code execution.
[PRIVILEGE_ESCALATION]: The provided code snippets for deserialization attacks and tool-use exploitation directly facilitate unauthorized access to system resources and sensitive data, representing a privilege escalation risk in the context of an agent's execution environment.

ai-ml-ctf