harvest
Warn
Audited by Gen Agent Trust Hub on Apr 30, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The helper script
scripts/generate-report.jsis vulnerable to command injection. It builds a command-line string for the GitHub CLI (gh) by directly concatenating the--repoand--authorarguments. This string is then executed usingchild_process.execSync, allowing shell metacharacters in the arguments to trigger arbitrary command execution. - Evidence: Inside
scripts/generate-report.js, the code usescmd += \-R \${options.repo}\;and thenexecSync(cmd, ...). - [COMMAND_EXECUTION]: The shell script
scripts/html-to-pdf.shcontains command injection vulnerabilities. It constructs command strings for headless Chrome andwkhtmltopdfusing input arguments and then executes them viabash -c. Because variables likeINPUT_FILEandOUTPUT_FILEare not properly sanitized before being passed to a sub-shell, malicious filenames could lead to arbitrary code execution. - Evidence: In
scripts/html-to-pdf.sh, the script defineslocal cmd=\"\\\"$chrome_cmd\\\" ... --print-to-pdf=\\\"$OUTPUT_FILE\\\" ...\\\"file://$INPUT_FILE\\\"\"and subsequently callsrun_with_timeout \"$cmd\", which performs abash -c \"$cmd\". - [PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection (Category 8).
- Ingestion points: The agent ingests untrusted data from GitHub PR titles, descriptions, and labels in the
COLLECTandANALYZEphases. - Boundary markers: No specific delimiters or instructions are used to ensure the agent ignores embedded instructions within the PR data.
- Capability inventory: The skill has powerful capabilities, including shell command execution via local scripts and file system manipulation.
- Sanitization: While the output scripts use
escapeHtmlfor display, there is no validation or filtering of the content before it is interpreted by the agent's reasoning engine, allowing malicious PR metadata to influence agent behavior. - [SAFE]: The skill's primary dependencies—the GitHub CLI (
gh),puppeteer, andchart.js(via a well-known CDN)—are reputable and appropriate for the skill's stated purpose of PR data visualization and reporting.
Audit Metadata