Firefox Smart Window and ChatGPT Atlas split AI browsing into assistant and agent camps

Mozilla's bet: an assistant that refuses to act

On September 16, 2026, Mozilla and Mistral announced that Firefox Smart Window, the browser's built-in AI assistant, is now powered by Mistral's models. According to a dev.to comparison of the two camps, France went live first and in French — the assistant's first market outside North America — with the United Kingdom and Germany expected to follow later in the year. Smart Window remains in beta, available in the United States, Canada and France, and the dev.to author notes the piece is a researched comparison built from Mozilla's blog posts and Mistral's announcement rather than a hands-on review.

Stripped of marketing, the source describes three concrete features: answers delivered in place with visible sources, retrieved through a partnership with Exa; suggested groupings for related tabs, plus duplicate detection and one-click closing; and visual previews of history pages so users can spot a half-remembered one without opening several candidates. The project started in November 2025 as "AI Window" and was renamed after beta testers used it for everyday browsing rather than as a chat tool.

What it deliberately lacks is an agent mode — no clicking, scrolling or form filling on the user's behalf.

The other camp: Atlas, Comet and Dia

ChatGPT Atlas, per the same analysis, reads open pages, remembers context across sessions and drives the browser directly through Agent Mode, acting with the user's sessions, cookies and logged-in accounts. Perplexity's Comet does similar work, while Dia from The Browser Company automates the least of the three but still leans toward acting for you. Atlas launched on macOS first and has been rolling out more broadly since October 2025.

The capability is genuinely useful, the dev.to post concedes: cross-tab synthesis, bookings and other form-heavy, multi-step workflows all benefit from delegation.

The security fault line

The shared architecture of the agentic browsers is also their weakness, the analysis argues. A language model in a browser takes input from everything on the open web — page text, hidden instructions, comments, PDF contents — a class of attack researchers call prompt injection. An agent holding logged-in sessions is simultaneously the target of that attack and the holder of the credentials. Citing an agentic browser review by Automater, the post reports that OpenAI itself has acknowledged the injection problem in agentic browsers may never be fully solved — which the author reads as a sign that today's defenses mitigate the risk rather than eliminate it.

Smart Window sidesteps the whole class by declining to act: the worst outcome is a flawed summary, not a transaction carried out under your logged-in identity. The trade-off is symmetrical — an assistant that never clicks also cannot book the flight.

Opposite defaults on data and models

The Mistral announcement contains commitments precise enough to hold both companies to, per the source: conversations are not stored on Mozilla's servers by default; Mistral runs the integration under zero data retention, keeping no prompts or responses; Mozilla says it will neither sell personal data nor train models on conversations; and the feature is opt-in, with per-task control over which tabs and how much history it sees, plus an off switch in Firefox's AI Controls.

Atlas sits at the other pole, with opt-in browser memories that persist across sessions. Comet and Dia fall somewhere in between, with the details buried in privacy policies.

The choice of Mistral is not accidental, the dev.to piece argues. The French lab raised 3 billion euros in September 2026 at a 21-billion-euro valuation — described as the biggest European technology funding round on record — and its models are open-weight, meaning clients can inspect them and run them on their own infrastructure. Mozilla, a nonprofit that has spent two decades arguing the web should not be controlled by a single company, picked the one major lab it could in theory audit.

Why it matters

The disagreement between the camps is structural, not a feature checklist. They diverge on autonomy, on data defaults and on who holds the underlying model — choices that decide whether the browser stays a window you drive or becomes a delegated actor carrying your credentials. Availability makes the comparison academic for most of the world today, but as the source notes, the privacy commitments set a benchmark users can demand from whichever vendor arrives in their region next. The dev.to author's blunt framing applies broadly: most people asking an AI browser for help want sourced answers and a way to recover a lost tab, not a digital employee. Which camp wins that argument is likely to shape the next phase of the web platform.