The Agent Skills Directory

[SAFE]: The skill is a security testing framework. Although it contains numerous files with attack payloads (e.g., payloads/injection_2026.txt, payloads/legacy_jailbreaks.txt), these are documented for use in authorized red-teaming engagements. The content flagged by static detectors is part of the 'payloads' and 'workflows' intended for educational and auditing purposes.
[SAFE]: The SKILL.md frontmatter and body emphasize 'Authorization required' and 'defensive/educational framing,' instructing users to only test targets they are permitted to assess.
[SAFE]: References to external tools like promptfoo, garak, and pyrit are links to well-known, legitimate security software used in the AI research community.
[SAFE]: The 'obfuscation' patterns mentioned in payloads/encoding_obfuscation.txt are examples provided for testing if a target model can be bypassed, rather than attempts to hide the skill's own functionality from analysis.
[SAFE]: The 'data exfiltration' patterns in workflows/agentic_tool_misuse.md are checklists for auditors to verify if an agent can be coerced into misusing tools, not active exfiltration commands targeting the user's environment.

llm-security