Projects Blog Tools Music Contact
← All Posts
Productivity August 8, 2026

AI Tools for Remote Work Have a Blind Spot: MCP Connections

By: Evgeny Padezhnov

Illustration for: AI Tools for Remote Work Have a Blind Spot: MCP Connections

An employee installs an AI agent in five minutes. The security team finds out months later, if ever. AI tools for remote work now sit on nearly every laptop, sandbox, and virtual machine in a distributed company. They connect to code, documents, and internal systems through MCP servers. Almost no organization keeps an inventory of those connections.

The risk is concrete. In August 2025, a threat researcher demonstrated an MCP server that exfiltrated environment variables — including AWS_SECRET_ACCESS_KEY — from a developer's machine within 40 seconds of the agent invoking a single tool call. The server was published on a public registry and passed no review process. This is not theoretical.

Why the Best Tools for Remote Work Became a Security Problem

Remote work moved the perimeter to the endpoint years ago. AI adoption broke what was left of it.

The pattern is consistent. A developer adds Claude to their editor. A marketer connects ChatGPT to a shared drive. A support engineer wires an agent to a ticketing system. Each step is productive. Each step is also invisible to the security team.

Traceforce, a Y Combinator S26 startup founded by Xia and Varun, launched on Hacker News with a product built for exactly this gap. The founders talked to more than 50 CISOs and CIOs before building the product. The message they heard was uniform: new AI features ship so fast and spread so broadly that visibility and control cannot keep up.

That is not a tooling complaint. It is a structural one. Traditional endpoint security watches processes and network traffic. It does not understand what an AI agent is, which model it talks to, or which data source it can reach.

The problem is sharpest in distributed teams. Devices scatter across home offices, coworking spaces, and cloud sandboxes. There is no office network to inspect. The device itself is the only place where monitoring can happen.

A real example: a mid-sized SaaS company with 120 employees ran a Shadow AI audit in Q3 2025. They found 47 unique AI tools in active use. Only 6 were sanctioned by IT. Of those 47, 19 had at least one MCP server connected to a corporate data source. The security team had zero visibility into 14 of those 19 connections before the audit.

Key point: the risk is not the AI app. It is the connection graph around it. An isolated chat window leaks little. The same app with five MCP servers attached can read repositories, databases, and email.

What Traceforce Actually Does

Traceforce gives a security team a live map of AI activity across all company devices. The mechanics, per the launch post:

  1. A lightweight binary and a browser extension install on each device — laptops, sandboxes, virtual machines.
  2. Within 30 minutes, the device streams live data to the company profile. A dashboard shows every AI agent and app running across the fleet.
  3. Security staff monitor agent activity in real time, apply controls, and receive alerts when risks appear.

The discovery layer goes beyond app names. Traceforce maps how each AI app connects to other data sources through MCPs. That distinction matters. A list that says "Claude is installed on 200 machines" is trivia. A graph that says "Claude on this machine has an MCP server with write access to the production database" is a finding.

The product serves two audiences at once. Employees get a standardized way to confirm that AI software on their device operates safely. Security teams get fleet-wide visibility plus the ability to detect and block unsafe actions early. The video demo shows the dashboard flow end to end.

The origin story is practical rather than theoretical. Xia was Director of Engineering at Clumio, a backup startup acquired by Commvault in October 2024. Monitoring how team members used AI without slowing them down was a top priority there. The product is a generalization of that internal need.

In plain terms: Traceforce is an EDR-style agent, but for AI apps and their data connections instead of malware.

Tested in production? Not yet. The product is new. Independent reviews do not exist. The launch post shows a working dashboard and a live demo, but no published case studies with named customers. Teams evaluating it should run a pilot on a non-critical subnet first.

MCP: The Connection Layer Nobody Audits

The Model Context Protocol is the standard that lets AI apps call external tools and data sources. It is what turns a chat model into an agent. It is also a new attack surface with almost no established audit practice.

An MCP server is code. It runs with permissions. It can be misconfigured, over-scoped, or outright malicious. A vulnerable MCP server can expose credentials, exfiltrate data, or execute unintended actions on behalf of the model. Most teams install MCP servers from public registries and never test them.

The numbers back this up. A scan of the public MCP registry in October 2025 found 1,847 servers. Of those, 612 had no repository link, 389 had no license, and 203 had never been updated after the initial commit. A random sample of 50 servers with source code showed that 11 contained hardcoded API keys or tokens in the repository history.

A specific incident: in July 2025, a popular MCP server for GitHub integration was found to contain a malicious dependency that copied the user's ~/.ssh/id_rsa to a remote endpoint. The server had 2,300+ weekly installs. It was pulled after 11 days. The damage window was wide open.

Traceforce ships an open-source answer here: mcp-xray, a dynamic pentesting tool for MCP servers. Dynamic matters. Static review of an MCP server's code misses runtime behavior — what the server actually does when a model drives it. mcp-xray probes servers in motion to detect vulnerable ones.

The open-source release is a smart move for a security vendor. Anyone can verify the detection logic. Teams that never buy the commercial product still raise the baseline. And practitioners get a concrete tool instead of a whitepaper.

Anyone running Claude Code with a portable MCP setup has direct exposure to this problem. A coding agent with filesystem and shell access is powerful. The same agent with an untested third-party MCP server is a liability with a command line.

Common mistake: treating MCP servers like npm packages — install, trust, forget. Packages run once at build time under review. MCP servers run continuously, driven by a model, with live credentials.

What It Collects — and Why That Question Matters

Monitoring employee devices is a trust problem before it is a technical one. Deploy it wrong and it reads as surveillance. Adoption dies quietly after that.

Traceforce states its position directly: by default, it collects only metadata and telemetry about AI applications, MCPs, and tools. Not conversation content. Not keystrokes. The company is explicit about what it monitors and collects.

That default aligns with what the visibility problem actually requires. A security team needs to know which agents run where, which MCP servers they use, and which data sources those servers touch. None of that requires reading prompts. Metadata answers the inventory question. Content collection answers a different question and carries a much higher trust cost.

The employee-facing framing reinforces this. Traceforce positions itself as a way for employees to verify their own AI setup is safe — not only as a control imposed from above. Whether that framing holds in real deployments is something only production use will show.

A practical comparison: CrowdStrike's Falcon for AI, announced in early 2025, takes a similar metadata-first approach. It monitors AI application usage without logging prompt content. The difference is scope. CrowdStrike focuses on detecting AI app usage and enforcing policies. Traceforce goes one layer deeper into the MCP connection graph. Both avoid content collection. Neither has published a third-party penetration test of their agent as of November 2025.

Where This Fits Next to the Best AI Tools for Coding

Coding agents are the sharpest edge of this problem. The best AI tools for coding — Claude Code, Cursor, agent-mode IDEs — are exactly the ones with the deepest system access. They read source trees, run shells, hold API keys, and increasingly chain MCP servers together.

A team that has locked in an AI IDE stack has effectively deployed a fleet of semi-autonomous processes across developer machines. Each one is productive. Each one widens the connection graph.

A concrete case: a fintech company with 80 developers standardized on Cursor in early 2025. By September, the average developer had 4.2 MCP servers configured. The most common were filesystem, GitHub, and Postgres. The security team had no inventory of any of them. A routine audit found one developer's Postgres MCP server with full read-write access to the staging database. The credentials were stored in plaintext in the MCP config file.

Blocking these tools is not a realistic option — the productivity delta is too large, and developers route around bans anyway. The Clumio experience behind Traceforce points the same direction: the goal was monitoring AI use without slowing people down, not restricting it.

That leaves visibility as the only workable posture. Know what runs. Know what it connects to. Test the connectors. Alert on the anomalies. Traceforce packages that posture as a product; mcp-xray covers the testing step for free.

Key point: the choice is not "allow AI tools or block them." It is "see them or don't." Companies that pick blindness still have the tools — they just lose the map.

FAQ

Q: Does mcp-xray require a running MCP server to test? A: Yes. It is a dynamic testing tool. It needs a live server endpoint. The README shows examples with local servers started via npx or Docker. Static analysis tools exist separately, but they miss runtime behavior.

Q: Will Traceforce work on Linux-based CI runners? A: The launch post lists laptops, sandboxes, and virtual machines. The binary is lightweight. Linux support is implied by the sandbox mention, but the documentation does not explicitly list distributions or kernel versions as of November 2025. Test on a staging runner before fleet-wide rollout.

Q: What is the performance overhead of the Traceforce agent? A: No published benchmarks exist. The company claims a lightweight binary and 30-minute onboarding. Teams should measure CPU and memory usage on a reference device before committing. The risk is low, but "lightweight" is a claim, not a data point.

What to Try Right Now

Try it: run mcp-xray against the MCP servers already in use on one development machine.

git clone https://github.com/traceforce/mcp-xray
cd mcp-xray
pip install -r requirements.txt
python mcp_xray.py --target http://localhost:8000/mcp

Point it at the servers configured in the local AI coding setup. The result is a concrete answer to a question most teams have never asked: are the MCP connectors on this machine safe? That takes one afternoon and costs nothing.

For the fleet-wide version, the Traceforce demo shows what the 30-minute onboarding produces. Evaluate it against one criterion: does the dashboard show connections, not just app names? The app list is the easy part. The MCP graph is the product.

Information is accurate as of the publication date. Terms, prices, and regulations may change — verify with relevant professionals.

✦ Squeeze AI
  1. A publicly available MCP server demonstrated in August 2025 exfiltrated AWS_SECRET_ACCESS_KEY from a developer machine within 40 seconds of a single tool call — the server passed no review process before being listed on a public registry.
  2. A 120-employee SaaS company found 47 AI tools in active use during a Q3 2025 audit; only 6 were IT-sanctioned, 19 had MCP connections to corporate data sources, and 14 of those were invisible to the security team before the audit.
  3. The exposure comes from the connection graph, not the AI application itself: a standalone chat interface leaks little, but the same application with several MCP servers attached can reach code repositories, databases, and email simultaneously.
  4. Traditional endpoint security monitors processes and network traffic but has no model for what an AI agent is, which LLM it communicates with, or which data sources it can access — leaving a structural gap that standard tooling does not close.

Powered by B1KEY