oma-scholar
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The file
resources/upstream-spec-cache.mdcontains directive language ([CLAUDE MUST DO]) written in Chinese, instructing the agent to proactively inform the user about API key configuration every time the skill is activated. While the file includes a disclaimer stating it is third-party reference material, the presence of imperative commands targeting agent behavior is a detected pattern. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from research papers (PDF, LaTeX, abstracts) to extract structured information. This creates a surface where malicious instructions embedded in paper text could attempt to influence the agent's output or actions during the sidecar generation process.
- Ingestion points: The skill ingests untrusted text from PDF files, LaTeX sources, and research abstracts.
- Boundary markers: While the skill enforces a strict YAML schema, it lacks specific boundary markers (e.g., delimiters) to separate processed untrusted text from internal instructions.
- Capability inventory: The skill has the capability to write local YAML files, perform network requests to academic APIs, and execute bundled Python scripts.
- Sanitization: The skill includes a
scripts/sanitize.pyutility intended to clean structural corruption in generated YAML, though its primary focus is on formatting rather than security sanitization of input content. - [EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch scholarly data from
knows.academy,api.openalex.org, andsemanticscholar.org. These are well-known academic services and are documented as the intended data sources for the skill's research functions. - [DYNAMIC_EXECUTION]: The skill executes bundled Python scripts, including
gen.py,lint.py,sanitize.py, andverify_metadata.py, to handle YAML generation, validation, and metadata verification. These scripts are run locally viapython3as part of the skill's core functionality.
Audit Metadata