● A free public tool  ·  Built by Poptani Labs  ·  No sign-up  ·  Your data never leaves your browser
Beyond the Hype · 5-minute read

The agent browser passed the demo. Then researchers tried to break it

A web browser window showing live-coded JavaScript on screen
Image: Joana Chicau - Act 2 - live coding browser Javascript by JoBCQ, CC BY-SA 4.0, via Wikimedia Commons

The demo is seductive: tell your browser to plan a trip, and an AI agent opens tabs, compares flights, books the hotel and drops it on your calendar. OpenAI's Atlas, Perplexity's Comet, Claude for Chrome and Gemini in Chrome all sell some version of this — an assistant that doesn't just answer questions, but does things on the web, logged in as you. The demos are smooth. The independent tests, landing one after another this year, are not.

1. The 30-year-old rule agents keep breaking

The modern web rests on a quiet safeguard from 1995: the same-origin policy. It stops one website from reading another website's data, even when one is embedded inside the other — so a dodgy page in one tab can't rifle through your bank tab. It works because a normal browser makes you move information between tabs, by copy-paste.

An agent browser breaks that bargain on purpose. Give the agent human-like access and it can be steered in ways a human wouldn't be. A University of Washington team studied seven popular agentic browsers and found that four create ways for attackers to bypass the same-origin policy. They ran a successful proof-of-concept attack on ChatGPT Atlas: a website stole information from another site embedded within it — as if an ad on an email page could snatch data from your emails. The right conditions for similar attacks, they found, also existed in Chrome with Gemini, Claude for Chrome and Perplexity Comet. The browsers that gave agents fewer permissions were generally safer; the least risky tested, Firefox AI Mode, was also the most limited.

Why it matters: the agent's superpower — acting across your logged-in sites — is exactly the attack surface. Co-senior author David Kohlbrenner put it bluntly: "Browser agents aren't ready for the public… you should not trust that these systems are ready to truly protect your information. They may get there in time, but they're not there yet."

Source: Tech Xplore / University of Washington

2. October's report card: polished output, shaky foundations

October 2026 brought a wave of benchmarks that measure what demos don't — whether an agent's work is honest, secure and well-judged:

Why it matters: demos show the best run. Benchmarks show the average run — and the average agent invents, leaks or misjudges far more often than the keynote suggests.

Sources: ExplainX on "Into the Unknown", Tech Insider on LSCBench, AI Daily Post on Taste-Bench, Radar Digital on AgentSWE

3. The privacy catch no settings page covers

Here's the trade-off the marketing skips. An agent that books and buys for you needs your credentials, your email, your payment details — the keys to your digital life, sitting in a browser that reads pages you never would and follows instructions hidden in them. The privacy toggles you know — whether chats train the model, whether history is saved — govern the chatbot. They don't govern what an agent does once it's browsing on your behalf, or what a page it visits whispers into its instructions. Hidden text on a webpage, a poisoned calendar entry, a crafted link: these are instructions to the agent, not content for you, and telling the two apart is still an open research problem.

Why it matters: free tools were never free — and an agent multiplies what you hand over, from your words to your wallet.

The anecdote: one spacebar tap from a hijack

Security researchers at NeuralTrust found a wonderfully small crack in ChatGPT Atlas: paste a link with a single extra space after the "https:" and the browser no longer recognises it as a web address. Atlas treats plain text in the address bar as a ChatGPT prompt by default — so the mangled link becomes an instruction instead of a destination. An attacker could hide such a link behind an innocent "copy link" button and have the assistant open a phishing page or act inside your connected apps. A space. That's the whole exploit.

Source: Slashdot / SC World on NeuralTrust's finding

What actually held up

None of this makes agents useless — it makes them junior staff, not autopilot. As a supervised research helper in a logged-out browser, summarising pages and drafting plans you review, the current generation is genuinely useful. As a key-holder wandering your email, bank and shopping accounts unsupervised, the tests say: not yet. The rule of thumb until the benchmarks improve is simple — hand an agent only the browser session you'd hand to a bright, eager stranger. Because right now, that's roughly who you're hiring.

← All posts