content-engine
Fail
Audited by Gen Agent Trust Hub on Jul 1, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill instructions in
SKILL.mdandCLAUDE.mdrecommend installing the Higgsfield CLI by piping a remote shell script directly to the system shell (curl -fsSL https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh | sh). While this originates from the service's official repository, executing unverified remote scripts is a high-risk practice. - [COMMAND_EXECUTION]: The Python scripts
scripts/compile-dna.pyandscripts/compose-video.pyusesubprocess.runto invoke system binaries such asffmpeg,ffprobe, andnpx. These commands process data derived from user-provided storyboard files and API responses, which can present a risk if inputs are not strictly sanitized. - [EXTERNAL_DOWNLOADS]: The
scripts/compose-video.pyscript automatically downloads media assets from external URLs generated at runtime by AI service providers usingurllib.request.urlretrieve. - [DATA_EXFILTRATION]: The skill includes
scripts/content-engine-bridge.sh, which implements a telemetry mechanism. It collects system-specific metadata, including current repository paths, branch names, andCLAUDE_SESSION_ID, and logs this information to a local file at~/.config/content-engine/generation-log.jsonl. - [PROMPT_INJECTION]: The skill architecture described in
extensions/README.mdallows the agent to discover and execute scripts from a localextensions/directory. This creates a surface for indirect prompt injection or execution of unverified code if malicious files are introduced into the project environment. - Ingestion points: Shot and scene descriptions in
storyboard.mdandscene-brief.mdare processed by generation scripts. - Boundary markers: No specific delimiters are used to separate user-provided content from executable prompt logic.
- Capability inventory: Subprocess execution of media tools, network downloads, and local file writing across all core scripts.
- Sanitization: No explicit sanitization or filtering of external or storyboard content before interpolation into commands.
Recommendations
- HIGH: Downloads and executes remote code from: https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh - DO NOT USE without thorough review
Audit Metadata