mac-code-local-ai-agent

Warn

Audited by Gen Agent Trust Hub on Mar 27, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The installation process involves cloning code from an external GitHub repository (github.com/walter-grace/mac-code) that is not identified as a trusted source.
  • [COMMAND_EXECUTION]: The agent architecture is designed to execute arbitrary shell commands via subprocess.run() based on LLM output. This capability represents a significant security risk if the agent's instructions are subverted.
  • [DATA_EXFILTRATION]: The skill facilitates data synchronization with Cloudflare R2, requiring the user to provide sensitive access keys (R2_ACCESS_KEY_ID, R2_SECRET_ACCESS_KEY). Users must ensure these credentials are not exposed through environment logging or shared sessions.
  • [PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection (Category 8) due to the combination of data ingestion and powerful execution capabilities.
  • Ingestion points: The agent ingests untrusted data from DuckDuckGo search results and local file content through the search and file tools.
  • Boundary markers: There are no explicit delimiters or specific prompt engineering instructions provided in the skill documentation to prevent the LLM from following commands embedded within the retrieved data.
  • Capability inventory: The system has the ability to execute shell commands, modify the file system, and perform network requests.
  • Sanitization: The provided instructions do not describe any mechanisms for validating, sanitizing, or requiring human approval for the shell commands generated by the model before execution.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 27, 2026, 01:41 AM