The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/run_experiment.py script executes evaluation commands (eval_cmd) using subprocess.run(shell=True). This poses a significant risk as it allows for arbitrary shell command execution, including command chaining and I/O redirection, if the experiment configuration is manipulated by the agent or an external actor.\n- [COMMAND_EXECUTION]: The scripts/setup_experiment.py script also utilizes subprocess.run(shell=True) to execute evaluation commands provided during the setup phase, creating a similar risk of shell-based exploits during the initialization of new experiments.\n- [INDIRECT_PROMPT_INJECTION]: The agent is instructed to read and act upon the contents of target files and a strategy document (program.md), which serves as an ingestion point for untrusted data.\n
Ingestion points: Target files and program.md strategy files read by the agent to determine optimization steps.\n
Boundary markers: The skill lacks delimiters or instructions to ignore potential commands embedded within the processed data files.\n
Capability inventory: The agent possesses the capability to execute shell commands (via the runner scripts) and perform git and filesystem operations.\n
Sanitization: No sanitization or validation is performed on the content of the files before they are ingested and processed by the agent.\n- [DYNAMIC_EXECUTION]: The core functionality of the skill is an autonomous "edit-and-run" loop where the agent dynamically modifies source code and then triggers its execution via a shell command. This provides a persistent mechanism for the execution of arbitrary, agent-modified code on the host system.

autoresearch-agent