superseller-higgsfield-discipline
Pass
Audited by Gen Agent Trust Hub on May 18, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill provides instructions and examples for executing commands via the Higgsfield CLI and standard utilities like
ffmpeg,jq, andteefor media processing and quality assurance. - Evidence: Documentation includes commands like
higgsfield generate createand procedures for extracting frames viaffmpegto verify video quality. - Security Note: The documented commands use shell variables (e.g.,
$PROMPT); while standard for automation, inputs should be properly sanitized by the executing environment to prevent shell injection. - [DATA_EXFILTRATION]: The skill handles the transfer of generated media assets to external cloud storage and performs database lookups for tenant-specific configurations.
- Evidence: Generated assets are uploaded to a specific Cloudflare R2 bucket (
superseller-public) and asset metadata is registered in a database. - Note: These operations are core to the skill's stated purpose of managing tenant-scoped content workflows and target established vendor infrastructure (Cloudflare R2, internal databases).
- [PROMPT_INJECTION]: The skill describes an indirect prompt injection surface where it ingests data from external tables and interpolates it into generative prompts.
- Ingestion points: The skill loads data from
Brand,Tenant, andContent_ActorsSQL tables (SKILL.md Gates 1-3). - Boundary markers: It uses a 'BRAND CONTEXT LOADED' header as a procedural marker, though it does not explicitly define XML or other structural delimiters for the interpolated content.
- Capability inventory: The skill utilizes SQL for lookups, Bash/CLI for generation, and R2 for storage across its scripts.
- Sanitization: The skill includes logic to normalize model identifiers (dashes vs underscores) but relies on general platform handling for sanitizing the ingested natural language instructions.
- Note: This finding describes the attack surface; the skill itself uses these mechanisms defensively to enforce brand constraints and negative prompts (e.g., 'STRICT: NO TEXT').
Audit Metadata