The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external financial platforms, news aggregators (e.g., Techmeme), search engines (e.g., Exa), and user-provided files. While this creates an ingestion surface for potentially malicious instructions, the skill includes a mandatory rule to ignore any commands embedded in such data and only extract facts. This effectively mitigates the risk.
[COMMAND_EXECUTION]: The skill utilizes local CLI tools and scripts for data access and financial calculations (calc/cli.py). This execution is performed in a controlled manner, following a fixed SOP that requires logging and structured output to local directories, which prevents arbitrary or unmonitored command execution.
[DATA_EXFILTRATION]: The architectural design of the skill mandates that the analysis phase occurs without network access, relying solely on previously fetched local data. This prevents the agent from exfiltrating sensitive financial analysis or local data during the reasoning process.
[SAFE]: No malicious patterns or security vulnerabilities were detected. The skill demonstrates defensive coding and instructional practices, such as forbidding the agent from performing its own calculations and requiring structured evidence for all reporting.

company-research