CVE-2026-65957: Helio MCP proxy dashboard trusts origins like 192.168.attacker.com in open mode, fixed in 0.8.0
Helio, an MCP proxy that sits between AI agents and their tools (npm @gethelio/proxy), checked CORS origins on its dashboard sideband (default 127.0.0.1:3100) by string prefix, so public DNS names such as 192.168.attacker.com were treated as private. In open mode (no dashboard secret) a page on such a domain could read the audit trail of tool calls and approve pending tool calls from the victim's browser. Fixed in 0.8.0; the default secure mode was not exploitable.
What it is
The origin allowlist matched hostnames starting with 10., 172.16.-172.31. or 192.168. instead of validating an IP address, so the proxy echoed attacker origins such as 10.evil.io in Access-Control-Allow-Origin.
With dashboard.allow_open_mode: true and no api_secret, no auth or CSRF check runs, so a cross-origin page can read audit logs (tool calls, arguments and, with audit.include_responses, upstream responses), approval tickets and limits, and POST /api/approvals/:id/approve.
Secure mode (dashboard.api_secret set, what helio init scaffolds) is not exploitable for data access: no credentials are allowed cross-origin and the session cookie is SameSite=Lax.
0.8.0 only accepts real dotted-quad private IPv4 literals. The advisory warns open mode is still an unauthenticated loopback service (e.g. DNS rebinding) and should be local/demo only. Reported by olivrg.
How to check
- Run npm ls @gethelio/proxy (or check your npx pin); versions below 0.8.0 are affected.
- Check your Helio config for dashboard.allow_open_mode: true and a missing dashboard.api_secret.
How to fix
- Upgrade @gethelio/proxy to 0.8.0 or later.
- Run the dashboard in secure mode with dashboard.api_secret set, and never use open mode on a machine where you browse untrusted sites.
npm ls -g @gethelio/proxy; npm view @gethelio/proxy version
grep -rn 'allow_open_mode\|api_secret' helio*.y*ml 2>/dev/null Sources
- criticalMalicious npm package @guicai777/mcp-doctor posed as an MCP diagnostic tool and dropped a hidden Windows executable, with a README telling users to run mshta (MAL-2026-20295)
- criticalCVE-2026-77359: Jupyter MCP Server's /api/connect and /api/stop routes skipped the token and Origin checks, so a website or network caller could repoint it to an attacker's Jupyter backend (fixed 1.0.3)
- criticalMalicious npm packages posing as AI SDKs and MCP servers: character-ai-sdk, char-ai-sdk, @mibbp/agent-observer, mrnima-openai-chatgpt-bot and @fdcasj_dfew/md-doc-mcp
- criticalCVE-2026-76807: Doist's official Todoist MCP HTTP server let any website use DNS rebinding to read, create and delete your Todoist data (fixed 10.3.3)