doc-speedreader
Warn
Audited by Gen Agent Trust Hub on Jul 22, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The script
scripts/convert.shis vulnerable to Python code injection due to unsafe string interpolation when callingpython3 -c. The variables$EXT,$URL, and$INPUTare placed directly into single-quoted Python strings without sanitization. If these inputs contain single quotes and additional Python commands, the interpreter will execute the injected code. - [COMMAND_EXECUTION]: The path validation logic in
scripts/convert.shfocuses on absolute path prefixes but does not account for path traversal techniques (e.g.,../../) that could be used to bypass directory restrictions. - [EXTERNAL_DOWNLOADS]: The
scripts/install.shscript downloads themarkitdownpackage and its various dependencies from the Python Package Index (PyPI) or a well-known mirror (Aliyun). While these are reputable services, the installation occurs dynamically at runtime. - [DATA_EXFILTRATION]: The skill allows the agent to fetch data from remote URLs to perform conversions. While it includes protections against Server-Side Request Forgery (SSRF) by blocking private IP ranges, it remains a capability that can be used to interact with external web services.
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted document content and passes it to the agent for summarization.
- Ingestion points: The
scripts/convert.shscript reads local files, data streams, and remote URLs provided by the user. - Boundary markers: The skill lacks explicit delimiters or instructions to treat document content as untrusted data, increasing the risk that the agent might follow instructions embedded within a processed file.
- Capability inventory: The skill possesses significant system capabilities, including shell script execution, Python code execution, file system writes (to the
task/directory), and network request capability. - Sanitization: No content-level sanitization or filtering is performed on the text extracted from the documents before it is processed by the agent.
- [SAFE]: The skill implements a security policy that explicitly blocks the agent from accessing sensitive system paths such as
/etc/,/root/, and credential-storing directories like~/.ssh/or~/.aws/.
Audit Metadata