neo-team-claude

Warn

Audited by Gen Agent Trust Hub on Apr 24, 2026

Risk Level: MEDIUMCREDENTIALS_UNSAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The system analyzer specialist is instructed to read and utilize sensitive environment files (e.g., .env, .env.sit, .env.uat) and Kubernetes secret objects (externalsecret) to facilitate infrastructure diagnosis.
  • [COMMAND_EXECUTION]: Employs a suite of powerful CLI tools including kubectl, psql, docker, and argocd for system investigation, alongside executing project-specific build and test scripts.
  • [REMOTE_CODE_EXECUTION]: The skill dynamically generates E2E test code files in TypeScript based on project contracts and then executes them at runtime using the system's test runner.
  • [EXTERNAL_DOWNLOADS]: Fetches and installs standard testing and development dependencies from the NPM registry during the initialization of the E2E test project environment.
  • [PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests untrusted user requests and local project context (e.g., CLAUDE.md, AGENTS.md) and interpolates them into sub-agent instructions. Ingestion points: User task descriptions and local repository documentation files. Boundary markers: Uses specific section headers to separate context but lacks explicit 'ignore embedded instructions' warnings for processed data. Capability inventory: Includes comprehensive file system access, shell command execution, and multi-agent orchestration. Sanitization: Relies on instructional 'Never Guess' constraints rather than structural data validation or escaping of external content.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 24, 2026, 01:53 AM