The Agent Skills Directory

[COMMAND_EXECUTION]: The instructions involve the use of several security and command-line tools such as hashcat, john, c-jwt-cracker, jwt_tool, and shell commands for Base64 manipulation.
[EXTERNAL_DOWNLOADS]: The skill recommends downloading external tools from GitHub (github.com/brendan-rius/c-jwt-cracker.git) and provides instructions for installing Python libraries such as pyjwt and cryptography via pip. It also uses the requests library to fetch content from remote URLs such as JWKS endpoints.
[DATA_EXFILTRATION]: The methodology provides examples for reading sensitive system files and credentials, including /etc/passwd, /flag.txt, and ~/.ssh/id_rsa, and mentions sending data to external attacker-controlled servers (attacker.com).
[DYNAMIC_EXECUTION]: The reference files contain multiple Python scripts designed to be executed at runtime for tasks such as RSA key extraction, token generation using public keys as HMAC secrets, and brute-forcing JWT secrets.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data in the form of JWT tokens from external targets. Ingestion points: JWT tokens parsed in SKILL.md (Phase 1). Boundary markers: None present; the agent is instructed to decode and act on claims without validation. Capability inventory: Includes file writing (open().write()), network requests (requests.get), and shell command execution via referenced tools. Sanitization: No sanitization or validation of the ingested token content is mentioned before processing.
[PRIVILEGE_ESCALATION]: The skill includes instructions to use chmod +x on a payload to make it executable as part of a command injection demonstration.

jwt-attack-methodology