The Agent Skills Directory

[PROMPT_INJECTION]: The instructions include explicit directives to bypass user oversight, such as "never stop, never ask permission" and "Run experiments indefinitely until manually interrupted." These patterns are designed to override the agent's standard safety protocols and user confirmation requirements when performing high-risk actions like code modification and shell execution.
[COMMAND_EXECUTION]: The skill instructs the agent to perform a series of shell operations using a local script (bash autoloop.sh). These commands are used to execute training runs, manage results, and commit changes to git. Because the agent is also instructed to modify the Python code that these shell commands execute, it creates a pipeline for autonomous, arbitrary code execution.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by reading local state such as git log, results.tsv, and run.log to determine its next action. While these are local files, they lack boundary markers or sanitization logic, meaning malicious data embedded in logs or results could theoretically influence the agent's code generation logic in the next iteration of the loop.

autoresearch