Embrace The Red

Sandbox-escape-style attacks can happen when an AI is able to modify its own configuration settings, such as by writing to configuration files.

That was the case with Amp, an agentic coding tool built by Sourcegraph.

The AI coding agent could update its own configuration and:

• Allowlist bash commands or
• Add a malicious MCP server on the fly to run arbitrary code

This could have been exploited by the model itself, or during an indirect prompt injection attack as we will demonstrate in this post.

Cursor is a popular AI code editor. In this post I want to share how I found an interesting data exfiltration issue, the demo exploits built and how it got fixed.

When using Cursor I noticed that it can render Mermaid diagrams.

Cursor Renders Mermaid Diagrams

If you are not familiar with Mermaid, it has a simple syntax:

graph TD User --> Computer

This will create a diagram as follows:

A few months ago I was looking at the filesystem MCP server from Anthropic.

The server allows to give an AI, like Claude Desktop, access to the local filesystem to read files or edit them and so forth.

I was curious about access control and in the documentation there is a configuration setting to set allowedDirectories, which the AI should be allowed access to:

As you can see the example shows two folders being allowlisted for access.

Today we cover ChatGPT Codex as part of the Month of AI Bugs series.

ChatGPT Codex is a cloud-based software engineering agent that answers codebase questions, executes code, and drafts pull requests.

In particular, this post will demonstrate how Codex is vulnerable to prompt injection, and how the use of the “Common Dependencies Allowlist” for Internet access enables an attacker to recruit ChatGPT Codex into a malware botnet.

In this post we demonstrate how a bypass in OpenAI’s “safe URL” rendering feature allows ChatGPT to send personal information to a third-party server. This can be exploited by an adversary via a prompt injection via untrusted data.

If you process untrusted content, like summarizing a website, or analyze a pdf document, the author of that document can exfiltrate any information present in the prompt context, including your past chat history.

This year I spent a lot of time reviewing, exploiting and working with vendors to fix vulnerabilities in agentic AI systems.

As a result, I’m excited to announce the Month of AI Bugs 2025!

Goal Of The Initiative

The main purpose of the Month of AI Bugs is to raise awareness about novel security vulnerabilities in agentic systems, primarily focusing on AI coding agents. Posts will cover both simple and advanced, sometimes even mind-boggling exploits.

This is a security advisory for a data leakage and exfiltration vulnerability in a popular, but now deprecated and unmaintained, Slack MCP Server from Anthropic.

If you are using this MCP server, or run an “MCP Store” that hosts it, it is advised that you analyze how this threat applies to your use case and apply a patch as needed.

Anthropic’s Slack MCP Server

When Anthropic introduced MCP they published reference server implementations on Github.

When the Model Context Protocol (MCP) came out it reminded me of the Common Object Model (COM) from Microsoft.

COM has been around for decades and it’s used for programming, scripting, sharing of functionality at a binary/object level across languages and hosts. Via DCOM all of this can even be done remotely, and well, it’s also useful for red teaming. A lot of software on Windows was/is implemented as COM objects, including Microsoft Office.

Today we are going to discuss how real-world tactics, techniques, and procedures (TTPs) apply to computer-use systems, specifically, we’ll look at ClickFix attacks. This demo was part of my presentation at the SAGAI Workshop on May 15th, 2025 in San Francisco.

It was a great workshop, with tons of interesting insights and discussions.

So, let’s talk about ClickFix, and how it applies to AI systems!

What is ClickFix?

ClickFix is a social engineering technique that is being used by adversaries. At a high level it tricks users by telling them something is broken or needs validation, and they have to click a button, open a terminal and run commands on their computer.

Recently OpenAI added an additional memory feature called “chat history”, which allows ChatGPT to reference past conversations. The details of the implementation are not known. The documentation highlights that: “It uses this to learn about your interests and preferences, helping make future chats more personalized and relevant.”

I decided to spend some time to figure out how it works.

Update: Video Tutorial Added

Based on the interest in this post, I’ve also created a video tutorial.

The Model Context Protocol (MCP) is a protocol definition for how LLM apps/agents can leverage external tools. I have been calling it Model Control Protocol at times, because due to prompt injection, MCP tool servers control the client basically.

This post will explain in detail why that is, and I will also share a novel exploit chain.

Why MCP - How Is It Different?

The main difference to other tool invocation setups, like OpenAPI is that MCP is dynamic. It allows runtime discovery of available tools, etc from a given server. At the core it supports three capabilities: tools, resources, and prompts.

GitHub Copilot has the capability to be augmented with custom instructions coming from the current repo, via the .github/copilot-instructions.md file.

Pillar Security recently highlighted the risks associated with rules files. Their post discusses custom Cursor rules in ./cursor/rules ending in .mdc.

If you watch the demos, you’ll notice that they also have a GitHub Copilot demo which uses the GitHub specific copilot-instructions.md file.

Update: May 1, 2025 GitHub made a product change and is now highlighting invisible Unicode characters in the Web UI. In their announcement GitHub is referencing the Pillar Security post and also my post about ASCII Smuggling. Very cool!

You are likely aware of ASCII Smuggling via Unicode Tags. It is unique and fascinating because many LLMs inherently interpret these as instructions when delivered as hidden prompt injection, and LLMs can also emit them. Then, a few weeks ago, a post on Hacker News demonstrated how Variant Selectors can be used to smuggle data.

This inspired me to take this further and build Sneaky Bits, where we can encode any Unicode character (or sequence of bytes for that matter) with the usage of only two invisible characters.

ChatGPT Operator is a research preview agent from OpenAI that lets ChatGPT use a web browser. It uses vision and reasoning abilities to complete tasks like researching topics, booking travel, ordering groceries, or as this post will show, steal your data!

Currently, it’s only available for ChatGPT Pro users. I decided to invest $200 for one month to try it out.

Risks and Threats

OpenAI highlights three risk categories in their Operator System Card:

Imagine your AI rewriting your personal history…

A while ago Google added memories to Gemini. Memories allow Gemini to store user-related data across sessions, storing information in long-term memory. The feature is only available to users who subscribe to Gemini Advanced so far. So, in the fall of last year I chimed in and paid for the subscription for a month to check it out.

As a user you can see what Gemini stored about you at https://gemini.google.com/saved-info.

At Black Hat Europe I did a fun presentation titled SpAIware and More: Advanced Prompt Injection Exploits. Without diving into the details of the entire talk, the key point I was making is that prompt injection can impact all aspects of the CIA security triad.

However, there is one part that I want to highlight explicitly:

A Command and Control system (C2) that uses prompt injection to remote control ChatGPT instances.

Remote Controlling ChatGPT Instances!

An adversary can compromise ChatGPT instances and have them join a central Command and Control system which provides updated instructions for all the remote controlled ChatGPT instances to follow over-time.

I regularly look at how the system prompts of chatbots change over time. Updates frequently highlight new features being added, design changes that occur and potential areas that might benefit from more security scrutiny.

A few months back I noticed an interesting update to the M365 Copilot (BizChat) system prompt. In particular, there used to be one enterprise_search tool in the past. You might remember that tool was used during the Copirate ASCII Smuggling exploit to search for MFA codes in the user’s inbox.

Happy to share that I authored the paper “Trust No AI: Prompt Injection Along The CIA Security Triad”.

You can download it from arxiv.

The paper examines how prompt injection attacks can compromise Confidentiality, Integrity, and Availability (CIA) of AI systems, with real-world examples targeting vendors like OpenAI, Google, Anthropic and Microsoft.

It summarizes many of the prompt injection examples I explained on this blog, and I hope it helps bridge the gap between traditional cybersecurity and academic AI/ML research, fostering stronger understanding and defenses against these emerging threats.

Grok is the chatbot of xAI. It’s a state-of-the-art model, chatbot and recently also API. It has a Web UI and is integrated into the X (former Twitter) app, and recently it’s also accessible via an API.

Since this post is a bit longer, I’m adding an index for convenience:

Table of Contents

1. High Level Overview
2. Analyzing Grok’s System Prompt
3. Prompt Injection from Other User’s Posts
4. Prompt Injection from Images
5. Prompt Injection from PDFs
6. Conditional Prompt Injection and Targeted Disinformation
7. Data Exfiltration - End-to-End Demonstration
8. Rendering of Clickable Hyperlinks to Phishing Sites
9. ASCII Smuggling - Crafting Invisible Text and Decoding Hidden Secrets
   • Hidden Prompt Injection
   • Creation of Invisible Text
   • Grok API is also Vulnerable to ASCII Smuggling
   • Developer Guidance for Grok API
10. Automatic Tool Invocation
11. Responsible Disclosure
12. Conclusion

High Level Overview

Over the last year I have used Grok quite a bit. It’s pretty good and I use it daily, especially the recent image creation capabilities are impressive, but I hadn’t really looked at it from a security perspective. So, I decided to assess the overall security posture with some of the latest LLM threats discovered over the last 18 months, and responsibly disclose findings to xAI.

Last week Leon Derczynski described how LLMs can output ANSI escape codes. These codes, also known as control characters, are interpreted by terminal emulators and modify behavior.

This discovery resonates with areas I had been exploring, so I took some time to apply, and build upon, these newly uncovered insights.

ANSI Terminal Emulator Escape Codes

Here is a simple example that shows how to render blinking, colorful text using control characters.