The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides explicit instructions to clone a repository and execute a Python script (smuggler.py) from an external GitHub account (defparam). This involves downloading third-party code and running it in the local environment without version pinning or integrity verification.
[EXTERNAL_DOWNLOADS]: The documentation encourages downloading and using external security utilities like unisub from unverified repositories such as github.com/tomnomnom/hacks.
[COMMAND_EXECUTION]: The skill contains several shell and PowerShell code snippets intended for generating attack payloads and testing injection points, which could be executed by an agent in an automated workflow.
[INDIRECT_PROMPT_INJECTION]: The skill methodology involves processing responses from untrusted external targets, creating a surface for indirect prompt injection if the agent ingests malicious target data.
Ingestion points: Web server responses and intercepted proxy traffic described in SKILL.md.
Boundary markers: No specific delimiters or safety instructions are provided to separate target data from the agent's instructions.
Capability inventory: Local shell execution (git, powershell) and remote script execution (python3) capabilities are referenced in SKILL.md.
Sanitization: The instructions do not detail any procedures to sanitize or validate data returned from external targets before it is processed by the agent.

payloads-ato