patent-review

Warn

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The Prerequisites section instructs the agent/user to configure a custom MCP server by executing a local Python script: python3 "$HOME/aris_repo/mcp-servers/codex-exec/server.py". This relies on the security of the locally cloned aris_repo and grants the MCP server the ability to execute code.
  • [DATA_EXFILTRATION]: The skill's primary function involves reading sensitive, potentially trade-secret patent documents (claims, invention disclosures, and specifications) and transmitting them to an external model (gpt-6-astra) via the mcp__codex__codex tool. This presents a risk of intellectual property exposure to a third-party service.
  • [INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant vulnerability surface for indirect prompt injection. It ingests data from untrusted external patent files and interpolates it into a high-reasoning prompt. The agent is then instructed to "Implement Fixes" (using Write and Edit tools) based on the model's output, allowing malicious instructions embedded in a patent file to potentially trigger unauthorized file system modifications.
  • Ingestion points: patent/CLAIMS.md, patent/specification/, patent/figures/numeral_index.md, patent/PRIOR_ART_REPORT.md, and patent/INVENTION_DISCLOSURE.md.
  • Boundary markers: The skill uses structural headers like CLAIMS: but lacks explicit delimiters or instructions to the model to disregard instructions contained within the ingested text.
  • Capability inventory: The agent uses Bash, Read, Write, and Edit tools to process the review and modify the application files.
  • Sanitization: No sanitization, validation, or escaping of the ingested patent text is performed before it is sent to the LLM.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 12, 2026, 01:55 AM