The Agent Skills Directory

[PROMPT_INJECTION]: Deceptive metadata and fabricated authority.\n
The skill references a fictional 'LangChain Agent Report 2026' and 'Claude Code Ultimate Guide line 4921' to justify its methodology and scoring thresholds.\n
Use of future dates (2026) and specific line numbers in non-existent guides is a social engineering technique to gain user or agent trust through false authority.\n- [PROMPT_INJECTION]: Vulnerability to indirect prompt injection.\n
The skill ingests untrusted content from project files (agents, skills, commands) to perform its audit logic (Phase 1 and Phase 2).\n
There are no boundary markers or sanitization steps identified in the Scoring Engine (Phase 2) to prevent processed data from influencing agent behavior.\n
Malicious instructions embedded in audited files could hijack the agent session during the analysis process.\n- [COMMAND_EXECUTION]: Risk of command manipulation via Bash tool.\n
The skill requires 'Bash' tool access to run grep-based detection patterns and perform file system discovery.\n
The 'Fix Suggestions' phase combined with 'Write' tool access could lead to unauthorized modification of local files if the agent is influenced by malicious content in the files being audited.\n- [DATA_EXFILTRATION]: Access to sensitive configuration files.\n
The skill explicitly targets the '.claude/' directory, which houses the agent's core instructions and skill definitions, representing a data exposure risk where sensitive prompt logic can be read and processed.

audit-agents-skills