r2-upload

Pass

Audited by Gen Agent Trust Hub on Nov 5, 2026

Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DATA_EXFILTRATION]: The r2_upload tool in src/index.ts allows the agent to read arbitrary local files using readFile(file_path) and upload them to an external S3/R2 bucket. This combined capability of filesystem reading and network writing serves as a potential exfiltration vector for sensitive data if the agent is instructed to upload files like ~/.ssh/id_rsa, .env, or other configuration files.
  • [INDIRECT_PROMPT_INJECTION]: The skill lacks built-in restrictions or sanitization for the file_path argument. It does not implement a blocklist for sensitive directories or files, making it vulnerable to indirect prompt injection attacks where an attacker-controlled input could trick the agent into reading and uploading sensitive user data.
  • [COMMAND_EXECUTION]: The skill instructions in README.md and SKILL.md require the user to execute several shell commands for setup and configuration, including pnpm install, pnpm run onboard, and pnpm run build. The onboard script (src/onboard.ts) is an interactive CLI that writes configuration data to the user's home directory (~/.r2-upload.yml). While these are standard developer setup procedures, they involve executing and building local code provided with the skill.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 5, 2026, 06:37 PM