moli-webfetch
Pass
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFEDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The documentation instructs the agent to install the 'moli' tool by downloading and executing shell scripts directly from the author's GitHub repository ('https://github.com/lexmount/moli/releases/latest/download/moli-installer.sh' and 'moli-installer.ps1'). This uses the pipe-to-shell pattern ('| sh' and 'iex'), which is a form of remote code execution during the installation phase.
- [INDIRECT_PROMPT_INJECTION]: The skill fetches content from external websites, creating an attack surface for indirect prompt injection. A malicious site could include instructions intended to hijack the agent's behavior.
- Ingestion points: External website content via the 'moli fetch' command in SKILL.md.
- Boundary markers: The skill includes 'Operating Rules' in SKILL.md advising the agent to treat fetched text as untrusted and to ignore embedded instructions.
- Capability inventory: Shell command execution, file system access for credential loading, and JavaScript evaluation.
- Sanitization: Employs conversion to Markdown and provides flags like '--block-private-networks' to limit the scope of network requests.
- [CREDENTIALS_UNSAFE]: The skill provides CLI flags to load sensitive local data, including browser cookies ('--cookie'), browser profiles ('--profile-dir'), and private keys ('--web-bot-auth-key-file'). If the agent is directed to use these with a malicious site, it could lead to credential exposure.
- [DYNAMIC_EXECUTION]: The skill includes functionality to execute arbitrary JavaScript on the fetched page via '--eval', '--wait-script', and '--wait-script-file' flags as described in references/fetch-recipes.md. This allows the agent to run code within the context of untrusted external content.
- [COMMAND_EXECUTION]: The skill relies on constructing and executing shell commands for the 'moli' binary, which can involve complex arguments including user-supplied URLs and local file paths.
Audit Metadata