higgsfield-troubleshoot
Fail
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute multiple local Python scripts (e.g.,
../../scripts/higgsfield_memory.py,scripts/seedance_lint.py) via the shell to log troubleshooting outcomes. These instructions include shell snippets that interpolate potentially unescaped user-influenced data, such as prompts and failure descriptions, directly into the command line (e.g.,python3 scripts/seedance_lint.py --confirmed "<prompt that passed>"). This pattern is vulnerable to command injection if the agent does not properly escape the arguments. - [PROMPT_INJECTION]: The text repeatedly references "HARD RULE 8" and "root SKILL.md" as authoritative constraints that must be followed. This usage of specialized, imperative language to override standard model behaviors (like prompt length) can be used to bypass platform instructions.
- [INDIRECT_PROMPT_INJECTION]: The skill establishes a data ingestion surface where user-provided prompts and performance feedback are stored in a local database (
../../db/). This creates an indirect injection risk if the stored content is later retrieved and processed by the agent or other scripts without strict sanitization.
Recommendations
- AI detected serious security threats
Audit Metadata