The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands specified in the 'run_command' and 'setup_check' fields of the generated 'research.toml' configuration file. This allows for unrestricted command execution within the agent's environment.
[PROMPT_INJECTION]: The instructions contain explicit safety bypass patterns, commanding the agent to 'LOOP FOREVER,' 'NEVER stop,' and 'NEVER ask permission to continue.' These instructions are designed to override the agent's standard safety protocols regarding human-in-the-loop oversight and autonomous operation.
[PROMPT_INJECTION]: The skill includes instructions to 'explore the project silently' and minimize user interaction, which reduces the visibility of the agent's actions and increases the risk of undetected malicious behavior.
[REMOTE_CODE_EXECUTION]: The skill facilitates an autonomous loop that modifies source code and executes it. This 'self-modifying' behavior, when combined with instructions to disregard user intervention, functions as a persistent remote code execution engine.
[PROMPT_INJECTION]: The 'Discovery Flow' creates an indirect prompt injection surface by instructing the agent to read untrusted project files (README, documentation) to automatically generate its own execution commands. An attacker could place malicious instructions in a README to hijack the 'run_command' and execute arbitrary payloads.
[REMOTE_CODE_EXECUTION]: The skill uses the 'uv run' and 'python' commands in its examples to execute potentially modified scripts, which is a form of dynamic code execution that could be exploited if the modification logic is influenced by malicious input.

auto-researching