The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in 'reference/agentic-simulation.md' direct the agent to execute shell commands for session cleanup (e.g., 'rm -rf "{session_dir}/"' and 'Remove-Item -LiteralPath "{session_dir}"') using a path constructed from the user-provided '[domain]' argument. Because these commands are executed in a shell environment and the 'domain' variable is not sanitized, a malicious user could provide input containing shell metacharacters or command substitutions (e.g., '$(command)' or '; command') to achieve arbitrary code execution on the host system. A similar risk exists in 'reference/simulation-evaluation.md' regarding the execution of 'unzip' on user-supplied file paths.- [PROMPT_INJECTION]: The skill ingests untrusted data from external Miro boards using the 'mcp__plugin_miro_miro__miro_list_board_items' tool during simulations and architectural analysis. This data is used to drive the agent's logic, making the skill vulnerable to indirect prompt injection where malicious board content could override agent instructions.\n
Ingestion points: Board items are read in 'SKILL.md' (via the '--discover-bcs' mode) and throughout the workshop phases described in 'reference/agentic-simulation.md'.\n
Boundary markers: None; the skill does not instruct the agent to wrap ingested board data in delimiters or to ignore embedded instructions within that data.\n
Capability inventory: The agent has access to shell execution ('rm -rf'), web research tools ('WebSearch', Perplexity MCP), and Miro board modification tools ('miro_create_sticky_note', etc.).\n
Sanitization: No sanitization, validation, or filtering of Miro board content is described before the agent processes and acts upon it.

simulation