github-kb
Fail
Audited by Gen Agent Trust Hub on Feb 15, 2026
Risk Level: HIGHPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- Indirect Prompt Injection (HIGH): The skill is designed to ingest and process untrusted data from arbitrary GitHub repositories. This presents a critical vulnerability surface.
- Ingestion points: The skill performs
git clone <repo-url>and subsequent file reads (Read,Glob,Grep) on the downloaded content (SKILL.md, README.md). - Boundary markers: There are no instructions to the agent to disregard instructions found within the cloned repositories or to treat the content as data-only.
- Capability inventory: The agent is instructed to use the
Explore agentvia theTask toolfor "comprehensive code analysis." This autonomous exploration likely includes the ability to process and act upon technical decisions and logic found in the code. - Sanitization: No sanitization or safety-filtering is performed on the content of the cloned repositories before analysis.
- Command Execution (MEDIUM): The skill explicitly instructs the agent to execute shell commands such as
git clone <repo-url>andcd {GITHUB_KB_ROOT}. While these are necessary for the skill's function, they provide a mechanism for path traversal or command injection if the repository URL or environment variables are manipulated. - External Downloads (MEDIUM): The skill facilitates the download of arbitrary code from external sources (GitHub). While the destination is categorized by the user, the source is untrusted and uncontrolled.
- Data Exposure (LOW): The skill operates within a user-defined directory (
GITHUB_KB_ROOT, default~/github-kb/). While it does not target sensitive system files by default, the agent's file-reading tools could be misused to access sensitive local data if the agent is successfully compromised via indirect injection.
Recommendations
- AI detected serious security threats
Audit Metadata