ghost-scan-deps

Fail

Audited by Gen Agent Trust Hub on Feb 20, 2026

Risk Level: CRITICALREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION] (CRITICAL): The skill performs piped remote execution by downloading a shell script from an untrusted external repository and immediately executing it.
  • Evidence: agents/init/agent.md contains the command curl -sfL https://raw.githubusercontent.com/ghostsecurity/wraith/main/scripts/install.sh | bash.
  • Source: The GitHub organization ghostsecurity is not in the trusted organizations list.
  • [EXTERNAL_DOWNLOADS] (HIGH): The skill downloads and executes a pre-compiled binary (wraith) from an unverified external source.
  • Evidence: agents/init/agent.md installs a binary to ~/.ghost/bin/wraith which is later executed by the scan agent.
  • [COMMAND_EXECUTION] (MEDIUM): The skill executes a downloaded binary with shell privileges to perform scanning operations.
  • Evidence: agents/scan/agent.md executes ~/.ghost/bin/wraith scan ... using the Bash tool.
  • [PROMPT_INJECTION] (LOW): The skill processes untrusted project files (lockfiles and source code) and feeds them into AI agents for analysis, creating an indirect prompt injection surface.
  • Ingestion points: Repository files (lockfiles and source code) entered via Glob and Read tools in agents/discover/agent.md and agents/scan/agent.md.
  • Boundary markers: Absent. No specific delimiters or instructions to ignore embedded commands are used when interpolating file content into agent prompts.
  • Capability inventory: The skill uses Bash (can execute any command), Task (can spawn subagents), and filesystem tools.
  • Sanitization: Absent. No evidence of content filtering or validation of ingested file data before it is analyzed by the subagents.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
CRITICAL
Analyzed
Feb 20, 2026, 02:04 PM