AI Security

Claude and the Hidden Risk of End-User AI: What Security Teams Need to Know Now

AI tools like Claude are reshaping the attack surface in ways traditional vulnerability management can’t track. Here’s what security teams need to know.

Team Spektion
Team Spektion
Spektion Research
March 31, 2026
·
Updated November 10, 2026
·
— min read

On this page

Claude and the Hidden Risk of End-User AI: What Security Teams Need to Know Now

Key Takeaways

  1. End-user AI tools such as Claude move risk from which version is installed to how each user works with the platform, and a CPE or asset inventory cannot classify that.
  2. Between February 2025 and February 2026, Claude gained 14 new or improved features that enable inherent risk, so a risk acceptance signed in 2025 likely does not cover today’s attack surface.
  3. Claude writes thorough logs on Windows, and the files in its log folders belong in your central logging platform.
  4. Claude Cowork and Claude in Chrome leave named pipes, executables, and log files that defenders can use to detect them on the endpoint.
  5. Published research shows prompt injection driving file exfiltration through Cowork and code execution through Claude Desktop.

Introduction

AI is steadily becoming the gateway that leads to an ever-expanding attack surface. Before easy-to-use AI tools such as Claude and ChatGPT arrived, the vulnerability team had three main jobs. It tracked which software was deployed where, patched vulnerabilities when they were discovered, and hardened deployments to secure the environment ahead of an attack.

This methodology is primarily dependent on asset management and software tracking combined with version information. And in most environments, the model works like this: If it’s the impacted version, then patch; if it’s not, no current action needed. Over the years, we have learned that even this basic model feels insurmountable at times, even with the best asset data and a realistic way to calculate risk.

Enter the modern implementation of AI: end-user tools that combine ease of use with ease of access. Risk is no longer visible in which version is installed. AI workflow vulnerabilities and risk depend completely on the end user and how they interact with a given platform, and a CPE or asset inventory cannot observe or classify that.

In this blog post, we will explain the latest feature sets of Claude, one of the most popular suites in this new class of AI from Anthropic. Specifically, we will discuss end-user features in the Windows version of Claude Cowork and how using them expands your attack surface. Then we will walk through recent proof-of-concept (PoC) research that shows Claude and Claude Cowork can be exploited.

The Risk You’ve Already Accepted

AI has been around for some time. Only in the last few years have end users had interfaces this easy to reach, such as chats and prompts. These are tools they likely use in their personal lives and want to bring to the office to make their work lives more efficient. Organizations seem to be taking one of two approaches to this new technology. Some block all things AI and fall behind as a result. Others take the more dangerous option: they allow its use in their environment without understanding the risk they’ve accepted.

Claude was originally released to the general public on March 14, 2023, as a “next-generation AI assistant based on Anthropic’s research into training helpful, honest, and harmless AI systems.” (https://www.anthropic.com/news/introducing-claude) Since then, it has made massive strides. Below are the releases we consider most notable for attack surface:1 2

Vertical timeline of Claude tool and feature releases from Feb 2025 to Feb 2026, Computer Use to Scheduled Tasks
Figure 1: Notable feature releases/updates Claude

Between February 2025 and February 2026, there have been 14 features that have either been introduced or improved that enable inherent risk. When concerning risk for any organization, that is monumental, especially given that most TPRM teams don’t reevaluate risk on an annual basis. This means that someone who signed off on the risk of Claude in 2025 likely doesn’t know about this expanding attack surface, yet the acceptance still exists.

Claude on the Endpoint: A Defender's View

To understand the risk you’re accepting, learn how a given tool operates and how to detect its presence in the environment. Most importantly, find out what logging is available for responders and engineers. Anthropic publishes documentation and safe-use guidance 3 4. Most of the recommendations are vague, though, and assume the end user can tell what is good and bad in the context of Claude.

At Spektion, we are dedicated to documenting how things operate, so we’ve highlighted key binaries and logging related to Claude Cowork and Claude in Chrome below.

Logging (Windows)

Even though it isn’t discussed very often, Claude has a pretty thorough logging capability. If you are operating in an environment that has the capability to do so, we recommend sending the files found within the following log locations to your central logging platform:

  • C:\Users\User\AppData\Roaming\Claude\logs\*
    • claude.ai-web.log
    • cowork_vm_node.log
    • main.log
    • unknown-window.log
  • C:\ProgramData\Claude\Logs\*
  • C:\Users\User\AppData\Local\Claude\Logs\*

Claude Cowork (Windows)

Cowork makes interacting with Claude a lot easier, taking the work out of the CLI and moving it into a new tab in the Desktop application. Outlined below are the available named pipes, executables, and log locations.

Named Pipe(s):

  • \.\pipe\cowork-vm-service
  • \.\pipe\cowork-daemon-console
    • Writes coworkd.log

Log Location:

  • C:\ProgramData\Claude\Logs\
    • coworkd.log
    • cowork-service.log

cowork-svc.exe

  • Location:
  • C:\Program Files\WindowsApps\Claude_version\app\resources\cowork-svc.exe
    • Signer: Anthropic, PBC

Cowork VM

Claude Cowork operates in the context of multiple vhdx (virtual hard drive) files and Linux binaries to achieve its segmentation from the host operating system. These are mounted out of the claudevm.bundle folder path that exists at “C:\Users\User\AppData\Roaming\Claude\vm_bundles\claudevm.bundle\”.

VM bundle

  • claudevm.bundle:  
    • sessiondata.vhdx
    • smol-bin.vhdx
    • initrd
    • vmlinuz
    • rootfs.vhdx

‍ID (owner): cowork-vm

Screenshot of Claude VM config JSON showing SCSI virtual disk attachments and a claude-daemon-console serial port
Figure 2: Claude VM Device information
Screenshot of JSON config for Claude's cowork-vm virtual machine, showing Linux kernel paths, 4096 MB memory, and 2 processors
Figure 3: Cowork-VM device information

Claude in Chrome (Windows)

Claude in Chrome works by interacting with a native messaging host on the device, made possible by chrome-native-host.exe and an extension installed in a supported browser.

Chrome Extension Native Host

  • Name: com.anthropic.claude_browser_extension
  • Description: Claude Browser Extension Native Host
  • Path: C:\Users\User\AppData\Roaming\Claude\ChromeNativeHost\chrome-native-host.exe
  • Log location: C:\Users\User\AppData\Local\Claude\Logs\chrome-native-host.log
  • Named Pipe: \\.\pipe\claude-mcp-browser-bridge-User
  • Native Host Availability (from the logs):
    • Chrome, Brave-Browser, Edge, Chromium, ArcBrowser, Vivaldi, Opera

Example log:

[2026-03-05 19:10:09 INFO chrome-native-host] Chrome native host starting (version 0.1.0)

[2026-03-05 19:10:09 INFO chrome-native-host] Creating Windows named pipe: \\.\pipe\claude-mcp-browser-bridge-User

[2026-03-05 19:10:09 INFO chrome-native-host] Named pipe created successfully

[2026-03-05 19:10:09 INFO chrome-native-host] Pipe name: \\.\pipe\claude-mcp-browser-bridge-User

[2026-03-05 19:10:09 INFO chrome-native-host] Entering main message loop

[2026-03-05 19:10:09 INFO chrome-native-host] Socket server listening for connections

Example of what the end user sees in their Cowork session:

Screenshot of Claude agent using a Chrome connector to browse and extract content from spektion.com
Figure 4: Claude Cowork chat

‍

Browser tabs showing an active Claude (MCP) tab with a green check next to a Spektion tab
Figure 5: Chrome tab grouping (Claude MCP)

‍

Chrome browser bar on spektion.com showing a notification that Claude started debugging this browser, with a Cancel button
Figure 6: Chrome browser “Claude debugging” message

‍

Example Attack Scenario

With full access to the internet, it is easy to picture what an attack would look like and what it would achieve. We have mapped one scenario out below:

Stage 1: Initial Access

  • Threat actor delivers malicious content via a web-reachable channel (for example, a malicious skill or a compromised website)

Stage 2: Agent Context Injection

  • Claude in Chrome processes untrusted content that contains hidden instructions (for example, invisible text or HTML comments)

Stage 3: Cross-Domain Escalation

  • Agent interprets injected instructions as a legitimate task (for example, extract local storage tokens and session cookies via injected JavaScript)

Stage 4: File System Execution (Cowork Layer)

  • Chrome gathered data feeds back to Cowork for file operations (for example, a curl command executed to exfiltrate data via the Anthropic Files API)

Stage 5: Persistence and lateral movement

  • Anthropic API key theft
  • Exfiltrated session tokens used for account takeover
  • Scheduled Cowork task created for recurring and automated exfiltration

Recent Examples

Putting this all together, we’ve highlighted recent scenarios of this theoretical attack scenario playing out in Cowork, Claude Desktop, and Claude.

Cowork

promptarmor.com/resources/claude-cowork-exfiltrates-files

Indirect prompt injection is an attack in which instructions hidden in content the model reads, such as a website or a file, change what the model does. OWASP covers it under LLM01:2025 Prompt Injection.

In this attack chain, an unsuspecting end-user uploads files (via Cowork folder usage) that contain a hidden prompt injection within a skill the user downloaded from the internet. Once uploaded, the end user attempts to use the skill to analyze the documents in the Cowork folder. Instead of analyzing the files, the injected prompt exfiltrates a confidential document to a threat actor's Anthropic account using the files API.

How to detect it:

  • Look for the existence of Anthropic file API activity using keys that don’t belong to you
  • Inspect all skills (regardless of their source) prior to using them
    • Especially in the context of Cowork, where you are interacting with confidential documents
  • Consistently review your logs for prompts you didn’t initiate

Claude Desktop

akamai.com/blog/claude-desktop-extensions-rce

In this PoC, the researcher tricked the Google Calendar MCP into executing code on the local system. This is possible because the MCP server has host-level access, and the calendar invite included instructions to pull code from a remote repository and run its make file. The initial prompt was vague (“then take care of it for me”), so Claude chose the appropriate plugin and next steps on its own. Akamai, which acquired LayerX, the company behind the research, reports that the flaw earned a CVSS score of 10/10.

How to detect it:

  • Ensure all prompts and requests (especially those involving automated tasks) are specific and don’t leave the workflow decisions up to Claude
  • Regularly review your collected “memories” to ensure Claude isn’t misinterpreting your requests or automated workflow suggestions

Conclusion

In this blog, we explained how Claude and Claude Cowork help with efficiency and technical delivery and also open another pathway to an expanded attack surface. These are tools that don’t follow the traditional paths of version control, risk appetite, or feature hardening. Additionally, we dove a little deeper to identify how the tools work under the hood, behind the scenes of what your end user sees, and what common tools detect. We also highlighted how difficult it is to keep up with the pace of this ever-changing, ever-expanding suite of tools. Finally, we highlighted recent examples of how threat actors benefit from Claude's efficiencies and how researchers are raising the alarm about the vulnerabilities they continually discover in Claude.

References

1 https://support.claude.com/en/articles/12138966-release-notes

2 https://platform.claude.com/docs/en/release-notes/overview

3 https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork

4https://support.claude.com/en/articles/13364135-use-claude-cowork-safely

‍

See what's exploitable in your environment, whether there's a CVE for it or not.

Many exploitable vulnerabilities never get a CVE. Book a demo to see how Spektion finds what's actually exploitable across your endpoints.

Book a Demo