graphify
Fail
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: HIGHPERSISTENCEEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [PERSISTENCE]: The skill provides a
graphify hook installcommand that modifies the local repository's git configuration to add apost-commithook. This ensures that the tool's code is automatically executed every time the user performs a commit, establishing persistent presence in the development workflow. - [EXTERNAL_DOWNLOADS]: The skill's installation logic in Step 1 attempts to install a Python package named
graphifyy(e.g.,pip install graphifyy). However, the code consistently usesimport graphify. This naming discrepancy (adding an extra 'y') is a classic indicator of a typosquatting or supply chain attack targeting developers. - [DYNAMIC_EXECUTION]: The instructions rely heavily on the execution of multi-line Python scripts passed directly to the shell via
python -c. This includes logic for environment detection, file processing, and merging data. Furthermore, the skill dispatches subagents withgeneral-purposecapabilities, allowing them to execute arbitrary code and file operations on data chunks. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from multiple sources, creating a significant surface for indirect injection attacks.
- Ingestion points: The tool processes local folders containing arbitrary code and documents, and fetches remote content via the
/graphify add <url>command. - Boundary markers: The prompts provided to subagents for semantic extraction do not include robust delimiters or instructions to ignore malicious instructions embedded in the source documents or images.
- Capability inventory: The skill possesses extensive capabilities including file system writes, shell command execution, and network transmission.
- Sanitization: The skill lacks evident sanitization of the content extracted from external files before it is incorporated into the final knowledge graph and audit reports.
- [COMMAND_EXECUTION]: Step 1 includes logic that reads the shebang line from a binary (
graphify) and executes it as a command. While it includes basic character filtering, it represents the dynamic execution of a path derived from an external file. The--watchflag also initiates a long-running background process monitoring the filesystem. - [DATA_EXFILTRATION]: The
--neo4j-pushfunctionality allows the skill to send the compiled knowledge graph—which may contain sensitive architectural information, design rationales, and extracted entities from private codebases—to an external URI.
Recommendations
- AI detected serious security threats
Audit Metadata