Security audit findings — tool description injection + missing output sanitization

Hi agent-infra team,

I recently ran a security audit on **sandbox** as part of research on MCP server security posture across the ecosystem.

Found a couple of items worth flagging:

**1. Tool description injection risk**
Tool descriptions in the sandbox execution environment aren't validated against adversarial patterns. Since this provides code execution capabilities to an LLM, a poisoned tool description could redirect the agent into executing malicious code, bypassing sandbox restrictions, or exfiltrating environment data.

**2. Missing output sanitization**
Code execution outputs and file system results returned to the model context aren't scanned for embedded injection patterns. A script's stdout could embed instructions that manipulate the agent's subsequent actions — especially dangerous in a sandboxed execution context where the LLM has broad capabilities.

Both are in a full audit report — 8-page PDF with CVSS ratings, EU AI Act mapping, and remediation steps — for $29 at luciferforge.github.io/mcp-security-audit/

Demo: https://luciferforge.github.io/mcp-audit-reports/

— Lucifer / LuciferForge Security