sd-real-engine
Warn
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: MEDIUMPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [PROMPT_INJECTION]: The skill contains a detailed 'compliance translation dictionary' (
references/impact-physics.md§五) designed to evade downstream platform filters. It explicitly lists prohibited violent terms (e.g., 'smashing eyes', 'crushing internal organs') and provides 'compliant' alternatives (e.g., 'muscle deformation', 'water splashes') to increase the probability of passing through AI model safety filters. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user requirements to generate complex prompts. While it mentions validation scripts (
validate_prompt.py), it lacks explicit boundary markers for user-provided data in the generation phase. This creates an attack surface where instructions embedded in user-provided scene descriptions could influence the final prompt output. - [METADATA_POISONING]: The skill claims to be the 'unique authority' for certain movement and combat rules and instructs the agent to ignore or override conflicting guidelines from other connected engines or the 'Master Control' (
short-drama-director). This internal override logic aims to ensure specific 'R1' behavior but can be used to bypass global safety guardrails established by the master controller.
Audit Metadata