The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple scripts, including autoresearch-agent/scripts/run_experiment.py and agenthub/scripts/result_ranker.py, utilize subprocess.run(shell=True) to execute commands defined in config.cfg or config.yaml files. This pattern is vulnerable to command injection if an attacker can modify the configuration files.
[DYNAMIC_EXECUTION]: The autoresearch-agent skill is designed as an autonomous loop that modifies source code and executes it to measure performance. This intentional capability allows for the dynamic generation and execution of arbitrary code within the agent's environment.
[INDIRECT_PROMPT_INJECTION]: Skills like agenthub and browser-automation read data from potentially untrusted sources (agent message boards, log files, and scraped web content) and interpolate it into the agent's context. The lack of explicit sanitization or strict boundary markers creates a surface for indirect prompt injection.
[DATA_EXFILTRATION]: The git-worktree-manager automatically copies sensitive files such as .env and .envrc into new worktree directories. While functional, this behavior increases the distribution of credentials across the file system.

engineering-advanced-skills