The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides functional code snippets and detailed methodologies for achieving Remote Code Execution through various bypass techniques.
Evidence in SKILL.md: Payloads for embedding PHP code in PNG metadata (IDAT/tEXt chunks) and using exiftool to inject command execution strings.
Evidence in SCENARIOS.md: Detailed exploitation flows for Tomcat PUT method (CVE-2017-12615) to upload JSP webshells and WebLogic (CVE-2018-2894) for arbitrary file upload.
[COMMAND_EXECUTION]: The skill includes specific shell commands intended to be executed on target systems to demonstrate compromise.
Evidence in SKILL.md: MVG (Magick Vector Graphics) and Ghostscript payloads containing shell commands like |id > /tmp/pwned for ImageTragick (CVE-2016-3714) exploitation.
Evidence in SKILL.md: Examples of webshell content such as <?php system($_GET["cmd"]); ?>.
[DATA_EXFILTRATION]: Instructions and payloads are provided for reading sensitive system files from the target's local file system.
Evidence in SKILL.md: FFmpeg HLS playlist payloads using the concat protocol to read /etc/passwd via file:///etc/passwd.
Evidence in SKILL.md: Payloads for SSRF targeting cloud metadata endpoints (e.g., 169.254.169.254) to exfiltrate IAM credentials.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a repository for malicious data (uploaded files) designed to influence or take control of downstream processing systems.
Ingestion points: File upload endpoints in target applications.
Boundary markers: None; the skill explicitly focuses on bypassing existing validation boundaries.
Capability inventory: The skill details capabilities for RCE, file system access, and cross-site scripting (XSS).
Sanitization: Instructions encourage finding paths where sanitization is absent or inconsistent across processing stages (Accept, Store, Process, Serve).

upload-insecure-files