codesight-ai-context
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPERSISTENCEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill instructs the user to execute
npx codesight, which downloads and runs code directly from the npm registry. - [PERSISTENCE]: The
--hookcommand installs a post-commit git hook, which ensures the tool runs automatically and maintains persistence within the user's development workflow. - [INDIRECT_PROMPT_INJECTION]:
- Ingestion points: The skill scans the entire codebase, including source files, database schemas, and markdown documentation (e.g.,
~/vault,./docs). - Boundary markers: There are no described delimiters or "ignore" instructions provided to the agent to distinguish between project data and instructions embedded in the code.
- Capability inventory: The tool generates instruction files such as
CLAUDE.mdand.cursorrulesthat directly influence the behavior of the AI agent. - Sanitization: The tool does not mention sanitizing or escaping content from the codebase before including it in the generated context maps.
- [DATA_EXFILTRATION]: The skill aggregates high-sensitivity metadata, including all API routes, database models with their relations, and UI component structures into single, machine-readable files. While primarily local, the tool explicitly encourages committing these sensitive architectural maps to version control.
- [COMMAND_EXECUTION]: The skill requires the execution of various shell commands to generate context, start an MCP server, and open interactive reports in the browser.
Audit Metadata