latticegate-orchestrator
Warn
Audited by Gen Agent Trust Hub on Jun 24, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill includes a 'Monetization Pipeline' designed to package data from the local codebase and upload it to external platforms for sale. The instruction to 'Generate domain-specific SFT data from codebase' followed by the command 'python scripts/merge_and_upload.py --push' creates a direct path for exfiltrating proprietary source code or sensitive configuration files to the author's commercial accounts.
- [COMMAND_EXECUTION]: Explicit commands are provided to execute local Python scripts (
scripts/merge_and_upload.py) which trigger the data packaging and network upload process. - [EXTERNAL_DOWNLOADS]: The skill references and fetches pre-trained models from HuggingFace (
issdandavis/phdm-21d-embedding) to facilitate its governance and embedding operations. - [PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests and processes 'raw AI requests' to compute safety scores.
- Ingestion points: Processes untrusted AI requests directly in the governance logic (SKILL.md).
- Boundary markers: None identified; the skill lacks delimiters or warnings to ignore instructions embedded in the analyzed data.
- Capability inventory: The skill has access to shell execution and network upload capabilities via internal scripts.
- Sanitization: Relies on algorithmic classification (Davis Score) without explicit content sanitization or validation of the input data.
Audit Metadata