The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation includes steps to download third-party security tools from GitHub repositories not identified as trusted vendors.
Evidence: git clone https://github.com/defparam/smuggler.git and git clone https://github.com/BishopFox/h2cSmuggler.git in SKILL.md.- [REMOTE_CODE_EXECUTION]: The workflow facilitates the execution of scripts downloaded from unverified remote sources.
Evidence: The commands python3 smuggler.py and python3 h2csmuggler.py are explicitly provided to be run after cloning the repositories.- [COMMAND_EXECUTION]: The skill uses local shell commands to perform fingerprinting and tool setup.
Evidence: Use of curl for architecture identification and git for dependency acquisition.- [INDIRECT_PROMPT_INJECTION]: The agent.py script possesses an attack surface for indirect prompt injection by processing untrusted network response headers.
Ingestion points: resp.headers in identify_architecture and response status line parsing in test_tecl_detection within scripts/agent.py.
Boundary markers: None present.
Capability inventory: The script can perform raw socket connections and write findings to a JSON file.
Sanitization: Untrusted header data is directly incorporated into the output report without validation.

exploiting-http-request-smuggling