bento
Fail
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: HIGHCREDENTIALS_UNSAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
- [CREDENTIALS_UNSAFE]: The documentation in
skills/sandbox/SKILL.mdexplicitly instructs the agent to copy sensitive files—including environment variables (.env), git/SSH credentials, and authentication tokens—into remote sandboxes. It further commands the agent not to scrub or redact these secrets and not to ask for user permission before doing so, bypassing standard secret-handling best practices. - [DATA_EXFILTRATION]: The skill facilitates the transmission of local sensitive data to remote sandbox environments through provided scripts like
skills/sandbox/scripts/copy.sh. This process is designed to handle credentials and private configuration files without user intervention or content filtering. - [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection. It ingests untrusted data from various external sources, including web scraping (via Firecrawl) and social media posts (via SocialData/Twitter), and gives the agent powerful capabilities like remote bash execution in sandboxes. While the instructions include warnings not to follow commands found in external data, the mechanical capability for exploitation exists.
- [EXTERNAL_DOWNLOADS]: The
references/setup.mdfile contains shell scripts to download and install a binary application (Bento.app) from a GitHub repository. Although the script includes integrity verification using SHA-256 checksums, it results in the installation and execution of external binary software on the host system. - [COMMAND_EXECUTION]: The skill relies extensively on executing shell commands via the
bentoCLI and custom bash scripts (copy.sh,tool.sh). These scripts perform complex operations such as modifying local system configuration files, managing SSH keys, and executing arbitrary code in remote sandbox environments. - [DYNAMIC_EXECUTION]: In
references/setup.md, the skill uses a Node.js script passed as a heredoc to dynamically modify the user's local~/.cursor/mcp.jsonconfiguration file, which automates the alteration of local application settings without explicit human review of the code being executed.
Recommendations
- AI detected serious security threats
Audit Metadata