The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the gemini CLI tool via a bash command using the --yolo flag. This flag is documented to auto-approve all tool calls requested by the Gemini model, bypassing user confirmation. This is high-risk as the model could be instructed to run arbitrary system commands if compromised by prompt injection.
[PROMPT_INJECTION]: User-provided arguments ($ARGUMENTS) are directly included in prompts sent to the Scientist, Mother, and Woman sub-agents. This creates a significant surface for prompt injection attacks, allowing a malicious user to override instructions or safety constraints.
[REMOTE_CODE_EXECUTION]: The combination of user-controlled input and the --yolo flag for tool call auto-approval effectively creates a path for remote code execution. An attacker could use prompt injection to induce the Gemini model to execute harmful tools or code on the local machine.
[DATA_EXFILTRATION]: The sub-agents are encouraged to read project files and have access to web search tools. A prompt injection could be used to read sensitive local data, such as environment variables or credentials, and then transmit that data externally via web search queries or command-line output.

magi-ex