The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/computer_use_runner.py utilizes subprocess.run with shell=True to execute commands. The trigger_fallback function takes a command string from a configuration object and executes it in the system shell.
[EXTERNAL_DOWNLOADS]: The skill is configured to download and add external components at runtime. It specifically references web-infra-dev/midscene-skills@computer-automation and am-will/codex-skills@gemini-computer-use which are fetched using the npx skills add command.
[DYNAMIC_EXECUTION]: The runner script manages dynamic context by executing shell commands to add new skills to the agent environment during an 'auto-recovery' flow when the primary vision AI fails.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates a significant surface for indirect prompt injection.
Ingestion points: The skill processes external research data files such as IBM SPSS (.sav), Stata (.dta), SAS (.sas7bdat), and Microsoft Excel (.xlsx) using various Python libraries. Additionally, Tier 3 uses Vision AI to interpret GUI screenshots of third-party software.
Boundary markers: There are no instructions or delimiters implemented to ensure the agent ignores potentially malicious instructions embedded within the statistical datasets or the software interfaces being automated.
Capability inventory: The skill has the ability to execute shell commands, perform COM automation for Windows applications, and write/execute syntax scripts (.sps, .prg, .do) for statistical software.
Sanitization: There is no evidence of input validation or sanitization for the content of the data files processed by the statistical stacks or automation scripts.

computer-use