The Agent Skills Directory

[COMMAND_EXECUTION]: The bash script bin/gemini-researcher is vulnerable to shell command injection. The variables $MODEL_ARG and $PASS_ARGS are expanded without quotes in the line gemini $MODEL_ARG $PASS_ARGS "$full_prompt". An attacker or a malicious prompt could provide input (e.g., for the --model flag) containing shell metacharacters like ;, &, or | to execute arbitrary commands on the host system.- [EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the @google/gemini-cli package from npm. This is an official tool from a trusted organization used for its intended purpose.- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it lacks safeguards when processing user data.
Ingestion points: User-supplied research topics, content, and URLs in bin/gemini-researcher.
Boundary markers: Absent; templates like DEEP_TEMPLATE and SUMMARY_TEMPLATE use direct interpolation without delimiters or instructions to ignore embedded commands.
Capability inventory: Subprocess execution via the gemini CLI.
Sanitization: Absent; the script uses printf and shell variable expansion to build prompts without escaping or validation.

gemini-researcher