The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses local debug logs at ~/.claude/debug/mcp-*.log which can contain sensitive conversation history, tool outputs, and identifiers. It provides a specific command to pipe log content to the gemini CLI for format validation, which involves transmitting potentially sensitive local data to an external AI service.
[COMMAND_EXECUTION]: Uses various shell commands such as tail, grep, curl, jq, lsof, ps, and node to monitor system logs, inspect running processes, and perform network requests to local MCP servers.
[PRIVILEGE_ESCALATION]: Recommends using the --dangerously-skip-permissions flag with the Claude CLI. This flag is designed to bypass standard user consent prompts for tool execution, which increases the risk of unauthorized actions being performed without user oversight during debugging.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes data from ~/.claude/debug/mcp-*.log. Ingestion points: local log files containing records of external server interactions. Boundary markers: None. Capability inventory: shell execution, network access via curl, and data transmission via gemini CLI. Sanitization: None. If an external MCP server provides a malicious response that gets logged, it could influence the agent's behavior when the skill subsequently reads and processes those logs.

mcp-spy