git-student-analysis

Pass

Audited by Gen Agent Trust Hub on Apr 5, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes local shell scripts (scripts/run_analysis.sh, scripts/extract_commits.sh) and Python scripts (scripts/analyze.py, scripts/generate_md.py, scripts/generate_xlsx.py) to extract metadata from Git history and generate reports. These operations are performed on the repository being analyzed and are essential to the skill's stated purpose.\n- [EXTERNAL_DOWNLOADS]: The skill clones Git repositories from user-provided URLs using git clone in SKILL.md. Additionally, scripts/setup_venv.sh downloads and installs well-known Python dependencies (pandas, openpyxl, python-dateutil) from the official PyPI registry to support data processing and Excel generation. All external resources are standard for the tool's functionality.\n- [PROMPT_INJECTION]: Indirect prompt injection surface identified through the processing of untrusted repository data.\n
  • Ingestion points: The skill ingests untrusted data from Git repository history, including commit messages and code diffs, which are fetched via git clone or read from a local path.\n
  • Boundary markers: The LLM analysis prompt defined in references/llm-diff-analysis.md uses structured sections (e.g., Diff: {diff_content}) and JSON response requirements, but lacks advanced delimiters to isolate the untrusted diff content from the system instructions.\n
  • Capability inventory: The agent has capabilities to execute local scripts, read files from the analyzed repository, and write report files to the local directory.\n
  • Sanitization: There is no evidence of sanitization or escaping of commit messages or code diffs before they are interpolated into the LLM prompts for qualitative analysis.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 5, 2026, 08:37 AM