The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides explicit instructions to bypass security controls. In the 'Common Pitfalls' section of SKILL.md (point 12), it instructs the agent to write code to a separate file (e.g., /tmp/fetch_latest.py) and execute it via the interpreter to 'avoid triggering security scanning approval prompt' that would otherwise occur with inline script execution.
[COMMAND_EXECUTION]: The skill directs the execution of various shell commands including package managers (uv pip install, pip install), file retrieval (wget), and headless browser operations (chromium --headless). It also references specific local file system paths like /home/ubuntu/ and ~/.hermes/skills/ for output storage.
[EXTERNAL_DOWNLOADS]: The skill fetches financial data from sec.gov. These are downloads from a well-known government service, which are integral to the skill's purpose, but involve downloading and processing external files at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from SEC filings (HTML, XML, and Markdown). It has an ingestion surface where external filing content is parsed using regex and then used in analysis, presenting a surface for indirect prompt injection if a filing contains maliciously crafted text.
[DYNAMIC_EXECUTION]: The skill facilitates the creation and execution of Python scripts at runtime. By generating scripts in /tmp/ and then executing them to perform data fetching and processing, it implements a dynamic execution pattern used to circumvent manual review checkpoints.

sec-filings