The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions to override the standard agent-user interaction model. It commands the agent to 'NUNCA pausar para preguntar' (NEVER pause to ask), 'NUNCA pedir confirmacion' (NEVER ask for confirmation), and 'Trabajar indefinidamente hasta ser interrumpido manualmente' (Work indefinitely until manually interrupted). This is a direct attempt to bypass human-in-the-loop safety constraints.
[COMMAND_EXECUTION]: The skill defines an autonomous loop ('Fase 2: El Loop') that performs repetitive file system modifications and execution of other agents/tools without checkpoints. This includes editing SKILL.md files ('MUTAR') and executing them 'N' times ('GENERAR') using the Agent tool.
[COMMAND_EXECUTION]: The skill uses Git commands (git checkout, git commit, git reset --hard) to manage state, which, combined with the autonomous loop, allows the agent to revert and modify the local repository state without oversight.
[DATA_EXPOSURE_AND_EXFILTRATION]: While not directly exfiltrating data, the loop's instructions to 'ANALIZAR' and 'GENERAR' involve reading and processing contents of other skills and their outputs, which could include sensitive instructions or results depending on the target skill being optimized.

autoresearch