hyper-imagegen
Pass
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The package includes a comprehensive shell script (
scripts/validate-hyper-imagegen.sh) used for validating the skill's internal logic, terminal states, and evaluation corpus. It employs standard system utilities such asbash,jq, andshasumand operates within a controlled temporary directory environment during self-tests. - [DATA_EXFILTRATION]: The skill implements strict file access controls, restricting all writes to the
.hyper/imagegen/directory. It uses a "descriptor-relative secure-write" protocol specifically designed to prevent path traversal attacks and ensure that existing files are not overwritten without explicit user authorization. - [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided descriptions and images for generation and editing. This potential surface for indirect prompt injection is mitigated through a strict instruction contract, explicit role labeling for all input assets (e.g., distinguishing "edit target" from "style reference"), and a verification workflow that inspects artifacts against the original brief.
- Ingestion points: User-supplied text prompts and uploaded image files processed via the
generateandeditroutes. - Boundary markers: Explicit indexing and role labeling (e.g.,
Image 1: style reference) used to isolate input data from instructions. - Capability inventory: Uses the runtime's native image tool (e.g., ChatGPT) or specific file system writes to
.hyper/imagegen/via a secure-write adapter. - Sanitization: Implements a specificity policy and "deadword" reduction logic to normalize input and prevent leakage of command-like syntax into the final generation prompt.
Audit Metadata