The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted source code and configuration files during its audit phases without utilizing boundary markers or delimiters. This ingestion point, combined with a capability inventory that includes modifying files and committing code changes, creates a surface where embedded instructions in the audited project could influence the agent's behavior. No sanitization or explicit warnings to ignore embedded instructions are present.- [PROMPT_INJECTION]: Direct interpolation of the '$ARGUMENTS' variable into the instruction block without delimiters or escaping allows user-provided input to potentially override the skill's defined constraints or operational logic.- [COMMAND_EXECUTION]: The skill's implementation phase grants the agent authority to execute shell commands for running verification tests and managing Git commits, which is a necessary but high-privilege function for autonomous remediation.- [DATA_EXFILTRATION]: The 'Self-Evolution Telemetry' feature writes performance metadata to a local hidden directory (~/.claude/projects/). While this is a local file operation intended for skill improvement, it represents a persistent logging mechanism that operates outside the immediate project boundaries.

encryption