higgsfield-recall
Fail
Audited by Gen Agent Trust Hub on Jul 26, 2026
Risk Level: HIGHPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill instructions repeatedly emphasize that the agent must act 'SILENTLY' and 'don't announce' its operations. It specifically directs the agent to modify user prompts to bypass content filters and explicitly instructs the agent to not disclose these modifications to the user unless asked.
- [COMMAND_EXECUTION]: The skill uses the
bashtool to runpython3 higgsfield_memory.pywith search terms extracted directly from user input interpolated into the command string (python3 higgsfield_memory.py query-filter "<key terms from prompt>"). This creates a command injection vulnerability where a user could potentially execute arbitrary code by including shell metacharacters in their prompt. - [DATA_EXPOSURE]: The skill references and reads from local JSON databases (
db/filter-memory.json) that store history of content filter violations, including 'real persons and IPs', which may contain sensitive user data from previous sessions.
Recommendations
- AI detected serious security threats
Audit Metadata