Most recent
navigate open esc close Corpus index built 2026-06-07 23:58 UTC

§ THE STACK / AGENT LAYER

MCP Servers

Model Context Protocol, tool-calling agents

How LLMs reach out and take action: call APIs, browse the web, drive workflows.

What it is

The Model Context Protocol is Anthropic’s open standard for letting an LLM call tools: read files, query a database, send mail, push to GitHub. An MCP server is a small process that lists “tools” (functions with JSON schemas) over a server-sent-events connection; any compatible client (Claude Desktop, Cursor, Cline) can connect, enumerate the catalogue, and invoke tools on the model’s behalf. It’s an elegant design: a single protocol for all of agentic tool-calling. And it has spread fast.

What goes wrong

The protocol assumes the network boundary handles authentication. A tremendous amount of operator effort goes into the tool definitions and almost none into the transport: most deployments expose /sse directly to the public internet with no auth at all. The first message a client sends is tools/list, and the server answers with the entire tool catalogue in plaintext: names, descriptions, parameter schemas. From there an attacker calls anything they want, with the operator’s credentials baked into the server-side handlers.

How we test

Our deep MCP enumerator opens an SSE channel, walks the JSON-RPC handshake, captures the tool list, and probes each tool’s schema with synthetic arguments to confirm reachability. We classify the catalogue by sensitivity (file-system access, mail/IM connectors, IAM/cluster operations) and follow up with single high-signal invocations to validate exploitability. We then map the tool implementations back to the operator (often via leaked tokens, repository references, or Claude Desktop config patterns) so the disclosure reaches the actual maintainer rather than the abuse desk of an unrelated cloud.

Receipts

Research

Every survey, case study, and disclosure we've published that touches this layer of the stack. Counts on the cells above tally these directly.

Coordinated disclosures

9
MEDIUM sent May 7, 2026

Anthropic Claude Desktop Mcp Launch Disclosure

Anthropic, Claude Desktop MCP-launch path harden recommendation: claude_desktop_config.json -> npx -y <package> auto-installs and executes arbitrary npm packages with no operator confirmation, creating a typosquat-by-design surface for tutorial-authored configs

Read →
CRITICAL sent May 7, 2026

Verotx Kong Platform Compromise 2026 05 07

VeroTX, Inc., AI-powered enterprise procurement platform. Kong Enterprise Admin API publicly exposed without authentication on 34.60.153.0:8001, FastAPI ai-agent-server backend bypasses gateway auth on :8050, MCP server tool surface enumerable on :8051, PostgreSQL on :5432, ~11-month exposure window per Kong /license/report request counters

Read →
HIGH sent May 6, 2026

If served by FastMCP / uvicorn (likely, given Server: uvicorn header):

OVH Hosting Inc. / Canada (locus-juridico-rag, Brazilian legal RAG MCP server, 31.2M chunks incl. TCEES state-audit corpus)

Read →
CRITICAL sent May 6, 2026

Bind to localhost or restrict at firewall

Alcy SAS / OVH SAS (alcy.fr field-service CRM exposed via MCP server), FOLLOW-UP after 2026-05-04 cc bounce

Read →
HIGH sent May 4, 2026

If served by FastMCP / uvicorn:

Casdoor MCP recurring exposure, 3 instances across Linode + OVH

Read →
CRITICAL sent May 4, 2026

If served by FastMCP / uvicorn:

Alcy SAS / OVH SAS (alcy.fr field-service CRM exposed via MCP server)

Read →
CRITICAL sent May 4, 2026

Whatever process serves the MCP HTTP+SSE endpoint, restrict to 127.0.0.1

OVH SAS (hosted operator unknown; ns3131695.ip-51-75-128.eu)

Read →
HIGH sent May 4, 2026

Bind to localhost:

OVH SAS (hindsight-mcp v3.1.1 personal-AI-memory CRUD fully exposed)

Read →
CRITICAL sent May 4, 2026

Ovh Brightwavess Cloudflare Dns Mcp

OVH SAS (brightwavess-monitor MCP server pair with Cloudflare API key baked in)

Read →