add-karpathy-llm-wiki

Warn

Audited by Gen Agent Trust Hub on Jun 22, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes various shell commands to create directories, modify project configuration files (using perl regex), and restart system services via launchctl (macOS) and systemctl (Linux).
  • [REMOTE_CODE_EXECUTION]: During setup, the skill runs local executable scripts including setup/index.ts via pnpm exec tsx and setup/lib/install-slug.sh using source. This allows for the execution of code logic stored outside the skill's primary instruction files.
  • [EXTERNAL_DOWNLOADS]: The skill defines a pattern for fetching external documents using curl to ensure full content ingestion for the wiki knowledge base.
  • [PROMPT_INJECTION]: The skill is designed to ingest and process arbitrary external data (URLs, PDFs, etc.) into a persistent wiki. It lacks clear sanitization or boundary markers to isolate source content from the agent's instructions, creating a surface for indirect prompt injection.
  • Ingestion points: sources/ directory and external URLs.
  • Boundary markers: None present in the instructions to the agent for processing source files.
  • Capability inventory: File system writes, shell command execution, service management, and dynamic skill generation.
  • Sanitization: No sanitization or validation of content from external sources is specified.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 22, 2026, 09:52 AM
Security Audit — agent-trust-hub — add-karpathy-llm-wiki