Security audit findings — tool description injection + missing output sanitization

[manja316]

Hi @Klavis-AI team,

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

Found a couple of items worth flagging:

1. Tool description injection risk
The platform's tool descriptions aren't validated against adversarial prompt patterns. As an integration platform running MCP tools at scale, a single poisoned tool description can propagate across all connected agents — making this a higher-severity issue than in single-server deployments.

2. Missing output sanitization
Tool outputs are forwarded to model context without scanning for embedded injection patterns. At scale, this means one malicious upstream API response can affect every agent consuming that tool.

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

— Lucifer / LuciferForge Security

[JiwaniZakir]

This report doesn't include any specific code paths, affected files, reproduction steps, or CVSSv3 vector strings — the technical details needed to actually triage and act on a security finding. The concerns described (tool description injection, unsanitized output forwarding) are real vulnerability classes in MCP ecosystems, but without identifying which specific handler, route, or middleware in the Klavis codebase is affected, there's nothing concrete to investigate here. Paywalling the actionable details of a reported vulnerability is not standard responsible disclosure practice — the norm is to share specifics privately with maintainers first, not to sell them. If there are genuine findings, the appropriate path is coordinated disclosure directly to the team with enough technical detail to reproduce and verify the issue.