graphify

Fail

Audited by Gen Agent Trust Hub on Oct 2, 2026

Risk Level: HIGHPERSISTENCECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [PERSISTENCE]: The skill provides dedicated commands to install a git post-commit hook (graphify hook install) and to integrate with a project's CLAUDE.md file (graphify claude install). The git hook is designed to automatically execute graph update logic every time a user performs a commit in the repository, allowing code associated with the skill to run persistently across different development sessions without explicit re-invocation.
  • [COMMAND_EXECUTION]: The skill relies extensively on shell command execution and inline Python scripting to manage its environment and process data. It dynamically resolves the Python interpreter (storing the path in graphify-out/.graphify_python) and uses it to execute complex logic blocks. This includes using pip install with the --break-system-packages flag to bypass safety restrictions in externally managed Python environments.
  • [EXTERNAL_DOWNLOADS]: The skill downloads and installs the graphifyy package from PyPI during its setup phase. It also supports cloning arbitrary GitHub repositories into a local directory (~/.graphify/repos/) when a URL is provided as an input corpus, which involves downloading and potentially executing external code.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest untrusted files (code, documents, images, video) and process them using LLM subagents for semantic extraction. This creates a potential vector where instructions hidden in the processed data could influence the subagent's behavior.
  • Ingestion points: Local files in the INPUT_PATH, remote URLs via graphify add, and external repositories cloned from GitHub.
  • Boundary markers: The subagent prompt in references/extraction-spec.md provides a strict JSON schema and formatting rules but lacks explicit 'ignore embedded instructions' markers or delimiters for the untrusted content being processed.
  • Capability inventory: Subagents are explicitly configured as general-purpose agents, granting them access to powerful tools like Bash (shell execution) and Write (file system modification).
  • Sanitization: The workflow relies on the subagent's instruction-following capabilities to only output JSON and write to specific temporary paths (graphify-out/.graphify_chunk_*.json).
  • [DYNAMIC_EXECUTION]: The skill generates and executes Python code at runtime by passing large, multi-line strings to the interpreter via bash command blocks. These scripts handle critical tasks such as merging extraction results, deduplicating graph nodes, and generating cumulative cost and analysis reports.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 2, 2026, 05:01 AM