ejentum-reasoning-harness

Warn

Audited by Gen Agent Trust Hub on Aug 10, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill setup requires the installation of the 'ejentum-mcp' package from the public npm registry, which is managed by a community author.
  • [REMOTE_CODE_EXECUTION]: The configuration uses 'npx -y ejentum-mcp' to download and execute code from an external source during the skill's operation.
  • [COMMAND_EXECUTION]: The MCP server setup involves executing shell commands ('npx') to run the tool bridge.
  • [PROMPT_INJECTION]: The skill is designed to retrieve 'scaffolds' (natural language instructions) from a remote API and directs the agent to 'ingest' and follow them, presenting a surface for indirect prompt injection.
  • Ingestion points: Remote instructions are brought into the agent's context through the output of the 'harness_reasoning', 'harness_code', 'harness_anti_deception', and 'harness_memory' tools.
  • Boundary markers: Absent; the instructions tell the agent to treat external scaffold fields as 'internal-reasoning instructions' without delimiters or safety warnings.
  • Capability inventory: The skill facilitates network operations (outbound HTTPS requests to 'ejentum.com') to fetch reasoning patterns.
  • Sanitization: There is no evidence of filtering or validation of the content received from the remote API before it is processed by the agent.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 10, 2026, 09:39 AM
Security Audit — agent-trust-hub — ejentum-reasoning-harness