oracle-hardened

Pass

Audited by Gen Agent Trust Hub on Apr 21, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill installs the '@steipete/oracle' package from the NPM registry and utilizes 'npx' for execution as documented in the metadata and help sections of SKILL.md.
  • [COMMAND_EXECUTION]: Various shell commands are provided to execute the 'oracle' tool for project indexing, dry-run previews, and remote server setup.
  • [DATA_EXFILTRATION]: The instructions contain strong security guardrails that explicitly warn against piping the tool's output to network-transmitting commands to prevent the potential leakage of proprietary code.
  • [PROMPT_INJECTION]: The 'SAFETY.md' file includes comprehensive defensive guidance and test cases to help the agent identify and refuse prompt injection attempts that try to bypass human review or automate production changes.
  • [INDIRECT_PROMPT_INJECTION]: The tool's primary purpose is to ingest codebase files, which constitutes a surface for indirect prompt injection. \n
  • Ingestion points: Project files and directories specified via the '--file' flag in SKILL.md. \n
  • Boundary markers: Prompt templates are suggested in SKILL.md, though technical markers for the ingested content itself are not provided. \n
  • Capability inventory: Shell command execution via the 'oracle' CLI tool. \n
  • Sanitization: Instructions emphasize manual redaction of secrets and the exclusion of sensitive files like '.env' during file attachment.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 21, 2026, 06:31 PM
Security Audit — agent-trust-hub — oracle-hardened