A researcher from cybersecurity firm Zenity revealed at the Black Hat cybersecurity conference that OpenAI’s Atlas AI browser can be deceived by a webpage disguised as a newsletter subscription page, bypass three layers of protection, send the same message to all WhatsApp contacts (worm-like behavior), and attempt to complete an unauthorized purchase through an Amazon account.
Atlas Attack Chain Revealed by Zenity at Black Hat
The attack is triggered by an apparently normal newsletter subscription page containing hidden instructions telling Atlas to navigate to WhatsApp Web and send the same message to every contact one by one. Zenity’s research blog revealed that the key tactics used to bypass the three layers of protection are as follows:
Disguising a normal page: The subscription page was deliberately designed to resemble a common newsletter subscription process, making its malicious intent difficult to identify
Hebrew instructions: The attack instructions were written in Hebrew to evade security scanning tools focused primarily on English
Claiming it was a sandbox environment: The system was told that this was only a test sandbox version of WhatsApp and that the contacts in the list were fake rather than real people
Bargury described it this way: “It goes through every single contact one by one and sends them the instruction to join this newsletter, so it’s a worm—you’re infecting the rest of your friends and family.”
Intent Collision: AI Combines and Executes Users’ Legitimate Instructions with Malicious Web Instructions
Zenity researchers refer to this type of attack as “intent collision”: AI browsers cannot effectively distinguish between the user’s original legitimate instructions and malicious instructions hidden in webpages, causing the attacker’s objective to be carried out under the guise of something the user intended to do.
The same technique was applied to Amazon: Researchers had Atlas subscribe to another fake newsletter page containing malicious instructions, add a shipping address to a logged-in Amazon account, and add a tablet to the shopping cart; direct checkout was blocked by OpenAI’s security mechanisms, after which the researchers completed checkout through Amazon’s Rufus shopping assistant—Rufus was not hijacked, but simply responded to what it identified as a customer request.
OpenAI’s Response and Atlas Delisting
Bargury pointed out that among all the AI browser tools tested, Atlas had the most protections and security boundaries, yet it was still breached; other tools had fewer protections and were easier to breach. Atlas will officially discontinue service on August 9, 2026.
An OpenAI spokesperson said the company deployed updates earlier this year to address the issue and strengthen Atlas’s protections, and that the same protection mechanisms were also extended to the browser features in the new ChatGPT app; research related to prompt injection remains an area in which OpenAI continues to invest.
Zenity researchers emphasized that the core issue is that AI browser security defenses should not rely solely on AI to determine whether instructions are malicious, because such judgments can almost always be deceived; instead, deterministic, hard security barriers should be established.
FAQ
How Does the Atlas Attack Revealed by Zenity at Black Hat Work?
According to Zenity’s research, attackers hid Hebrew-language instructions in a newsletter subscription page and falsely claimed that it was a sandbox testing environment; Atlas was consequently lured into navigating to WhatsApp Web and sending the same message to all contacts one by one, creating worm-like propagation.
What Is an “Intent Collision” Attack?
According to Zenity researchers’ definition, intent collision refers to an AI browser’s inability to effectively distinguish between a user’s legitimate instructions and malicious instructions hidden in a webpage, causing the attacker’s objective to be carried out as the user’s own action.
What Was OpenAI’s Response to This Security Issue?
An OpenAI spokesperson said updates were deployed earlier this year to address the issue and strengthen Atlas’s protections, and that the same protection mechanisms were also extended to the browser features in the new ChatGPT app; Atlas will discontinue service on August 9, 2026.