higgsfield-generate

Fail

Audited by Gen Agent Trust Hub on Jul 19, 2026

Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill's bootstrap section instructs the agent to install a CLI tool by downloading a shell script from a remote GitHub repository and piping it directly to a shell: curl -fsSL https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh | sh. This pattern is high-risk as the remote script is unverified and its content could be changed at any time to execute malicious code on the host system.
  • [COMMAND_EXECUTION]: The instructions make heavy use of the Bash tool to run the higgsfield CLI. Commands are constructed by interpolating user-provided data, such as prompts and URLs, directly into shell command strings (e.g., higgsfield generate create ... --prompt "..."). This lack of strict sanitization or the use of safe argument passing mechanisms (like the -- separator) increases the risk of command injection.
  • [DATA_EXFILTRATION]: The skill uses network-capable tools (curl, higgsfield CLI) to transmit user data, including prompts and media, to external servers. It also supports fetching data from arbitrary URLs provided by the user, which can be misused to probe internal network services or exfiltrate environment information.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its URL fetching functionality. When the agent fetches product information from an external URL, the retrieved data (titles, descriptions, etc.) could contain malicious instructions designed to hijack the agent's context or influence subsequent tool calls.
  • Ingestion points: External data enters the context via higgsfield marketing-studio products fetch --url <url> in SKILL.md.
  • Boundary markers: Absent; external inputs are interpolated into shell commands and agent prompts without clear delimiters.
  • Capability inventory: The skill uses Bash to execute the higgsfield CLI, curl, jq, and mktemp across its primary workflow.
  • Sanitization: No validation or escaping of the content retrieved from external URLs is performed before it is processed or presented to the agent.
Recommendations
  • HIGH: Downloads and executes remote code from: https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh - DO NOT USE without thorough review
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Jul 19, 2026, 05:17 PM
Security Audit — agent-trust-hub — higgsfield-generate