image-generation
Fail
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: HIGHDATA_EXFILTRATIONPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [DATA_EXFILTRATION]: The
generate_poster.tsscript allows reading arbitrary local files through the--assetsargument and transmits their contents to external AI provider endpoints (fal.ai and Google). The script does not validate the file type before reading and uploading the data. - [PROMPT_INJECTION]: Instructions in
SKILL.mdprovide snippets using high-priority keywords such as 'CRITICAL:' and 'MANDATORY:' to influence agent behavior and modify prompt generation logic. - [INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect injection by accepting user-defined prompts and file paths while possessing broad filesystem and network permissions. Evidence includes the use of
readFileSyncfor file access andfetch/uploadfunctions for network communication without sanitization or boundary markers between user data and system commands. - [COMMAND_EXECUTION]: The skill instructs the agent to run local scripts using
npx ts-node. The implementation ofcopyToDestinationingenerate_poster.tsuses the user-provideddestinationpath directly in acopyFileSyncoperation without sanitization, which could allow files to be written to arbitrary or sensitive system locations via path traversal.
Recommendations
- AI detected serious security threats
Audit Metadata