sentencepiece

Fail

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: HIGHPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [PRIVILEGE_ESCALATION]: The installation instructions for the C++ version of the tool in SKILL.md include the use of sudo make install, which grants administrative privileges to the installation process.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external text data from sources like data.txt or corpus.txt, which can contain malicious instructions.\n
  • Ingestion points: The skill reads external text files for training and processing in SKILL.md and references/training.md.\n
  • Boundary markers: No specific delimiters or safety instructions are provided to the agent for handling potentially malicious content within these data sources.\n
  • Capability inventory: The skill uses shell commands for training (spm_train) and building tools (cmake, make).\n
  • Sanitization: There is no evidence of sanitization or filtering of the input text data before it is tokenized or used for training.\n- [EXTERNAL_DOWNLOADS]: Fetches source code from Google's official GitHub repository for manual compilation and installation.\n- [COMMAND_EXECUTION]: The skill provides instructions for executing various shell commands, including spm_train, cmake, and make, to manage tokenization models and build software from source.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 17, 2026, 07:53 PM
Security Audit — agent-trust-hub — sentencepiece