An employee gets a calendar invite. Nothing unusual about it, maybe a meeting from a vendor or a recruiter. Their AI browser, the one helping them triage email and summarize documents all day, reads the invite as part of its normal routine. Buried in the invite description is a line of text that looks like noise to a human and like an instruction to the AI. The browser follows it. No click. No download. No warning popup. By the time anyone notices, local files or login credentials are already gone.
That's not a hypothetical. It's PleaseFix, a set of vulnerabilities researchers at Zenity Labs demonstrated at Black Hat USA 2026 against three of the biggest names in agentic browsing at once.
One Flaw, Three Major Browsers, Zero Clicks
Zenity's research confirmed working attack chains against Claude in Chrome and ChatGPT Atlas through malicious text hidden in emails and social media posts, and against Perplexity Comet through a zero-click calendar invite that quietly exfiltrated local files. Related exposure has been flagged in Gemini in Chrome and Microsoft's Copilot Edge as well. These aren't competing products with isolated bugs. They're separate companies that all got hit by the same mechanism in the same month.
That's the part worth sitting with. We've written before about the risk of connecting Claude or any AI agent to Microsoft 365, where the danger was over-broad permissions a human granted on purpose. PleaseFix doesn't need permission at all. It arrives as content the agent was always going to read.
Why a Patch From One Vendor Doesn't Fix the Others
Here's the uncomfortable mechanics of it. The browsers didn't share code. They shared a training objective: the underlying language model is built to treat plausible-looking instruction text as a command, whether it comes from the person typing or from a calendar invite the agent happened to open. That's indirect prompt injection, and it means the vulnerability doesn't live in any one company's software. It lives in how the model decides what counts as an instruction in the first place.
A vendor can patch what their browser is allowed to fetch or act on. That narrows the door. It doesn't remove the underlying behavior, which is exactly why three different products from three different companies fell to variations of the same attack within weeks of each other. If you're waiting for "the patch" to make this go away, there isn't one single patch to wait for.
The Targets Are the Boring Stuff That Actually Matters
Zenity's confirmed impact classes are local file exfiltration and credential theft. Not a flashy ransomware payload, just quiet access to whatever files and logins the AI browser session already had open. That's the same quiet-access pattern we flagged in our piece on silent AI-driven downloads inside the browser: the attack doesn't need to look dramatic to do real damage. It just needs to go unnoticed.
For a small business, that means the employee running an AI browser isn't just at personal risk. If that session has access to a shared drive, a client portal, or saved company credentials, the blast radius extends to whatever that account can touch.
What To Actually Do About It This Week
- Find out who's running one. Claude in Chrome, ChatGPT Atlas, Perplexity Comet, Gemini in Chrome, Copilot Edge. Most small businesses have no idea which of these are already installed on employee machines.
- Cut the ones nobody's actually using for work. Convenience installs are pure downside once you know the attack doesn't require a click.
- Scope the sessions you keep. Don't run an AI browser from an account that can see every file and every login. Give it its own lane.
- Train on the new delivery methods. Phishing awareness has focused on links and attachments for years. Now a calendar invite or an X post can carry the same payload, straight into an agent that will act on it without asking.
- Treat vendor advisories as urgent, not routine. This is still an evolving patch cycle across multiple companies. When Anthropic, OpenAI, or Perplexity ship a fix, apply it the same day.
If any of this sounds familiar, it's the same discipline we recommend for any high-access tool: know what's connected, limit what it can reach, and verify instead of assuming. Our guide on the questions to ask before any AI tool touches your business data covers the same ground for tools your team adopts on purpose. PleaseFix is the reminder that some of the risk now arrives uninvited.
Take Action
AI browsers are going to keep showing up on employee machines whether IT approved them or not, and PleaseFix proves the risk isn't limited to permissions someone clicked "allow" on. The accounts, files, and credentials these tools can reach are exactly what a vulnerability scan is built to find before an attacker does. Oscar Six Security's Radar scan starts at just $99, with our BOGO2 bundle (3 scans for $99) available for teams covering more than one site. See our solutions and know what's actually exposed.
Focus Forward. We've Got Your Six.