Skip to main content
Ernesto Amato
Ernesto AmatoFractional CAIO & AI Strategist
← Back to Insights
•Emerging AI

On the Agentic AI Browsers War, the End of the Web, and the Dawn of the Agenticverse

Ernesto Amato
Ernesto Amato
Executive Leader & Fractional Chief AI Officer

OpenAI’s ChatGPT Atlas was announced on October 21, 2025.

We were already familiar with Comet by Perplexity, and we know that Google offers Gemini within Chrome and Gemini Computer Use (the latter, for now, via API, awaiting the release of Gemini 3 Agent).

Then there are niche and specialized products like DIA (The Browser Company / formerly Arc) and Fellou. Currently, the only major player missing from the roster is the Edge/Copilot duo…

We hear a great deal about browsers equipped with integrated AI agents capable of performing actions and completing tasks on behalf of human users.

What Are Agentic AI Browsers?

Essentially, these are standard web browsers (almost all based on the open-source Chromium project) integrated with generative AI models capable of:

  • having read/write access to all pages visited by the browser and/or applications present on the user’s computer;
  • maintaining persistent memory of choices, preferences, and actions taken by the individual over time.

Some of these Agentic AI browsers require access to all available user information right from the initial installation phase (for example, Comet allows linking your Google account). In this way, they begin to “learn” user behaviors and habits from first configuration.

The promised benefits of this new category of web navigation tools are multifaceted. Because they interpret the semantic content of pages being browsed, the AI model understands context and can interact autonomously.

The Four Promised Application Areas

1. Interaction with Page Content

This ranges from automatic form completion (raising massive ethical questions regarding, for example, the autonomous completion of online academic exams) to traversing pages and services by clicking buttons and hyperlinks.

2. Automating Actions on Behalf of the User

Users can delegate entire checkout procedures on e-commerce sites (exemplified by recent debates between Perplexity and Amazon regarding Comet’s use for automated purchases) or booking tables at restaurants and travel itineraries. There is already growing momentum behind Agentic E-commerce.

3. Personalized Assistance

By granting access to online services like email, Reddit, Discord, or content management systems like WordPress, our agentic browser can execute complex operations on our behalf in a customized, proactive manner.

4. Enterprise Workflow Management

Any activity coordinated through cloud platforms falls within the reach of agentic browsers: social media management, marketing campaign scheduling, or analytics reporting.

Fellou goes further: it requests permissions to interact directly with other desktop applications and the underlying operating system itself, scaling operational automation to an unprecedented degree.


These are just a few examples of what AI browser agents can already accomplish, and new use cases emerge weekly.

However, I believe three critical areas demand rigorous scrutiny.

Three Critical Dilemmas We Must Address

1) Illicit Use and Fraud

A vast proportion of academic and professional certification programs are delivered online, featuring remote proctored exams.

Experiments have already been conducted in this space with disturbingly capable results. The immediate response from browser AI model developers has been to firmly condemn this usage, as stated by Perplexity CEO Arav Srinivas:

“Absolutely don’t do this!”

Yet the objective reality remains: this illicit usage is not technically prevented by the models themselves.

2) Privacy and Data Governance

The second consideration concerns the data we entrust to these agentic browsers. How is the immense volume of personal, private, and confidential data gathered in real-time actually handled?

Reviewing their privacy policies reveals significant reasons for concern.

Dia – The Browser Company

The Browser Company’s agentic browser, Dia, natively includes ad, tracker, and cookie banner blockers: a positive start.

The home page states: “Privacy in Dia, You are in Control”, linking to their data privacy section. Initial impressions appear favorable: data is encrypted and stored locally on the user’s device. Only prompts are transmitted to Dia servers and forwarded to unnamed “AI Partners”, with clauses stating data is not used for model training:

“These partners are restricted from training on your data, and may not store it after your request is complete.”

Dia also states it uses data with consent to improve services while attempting to avoid processing sensitive banking or healthcare sites.

However, a critical factor for European Union residents is that Dia is hosted and operated in the United States. Personal data is transferred to and stored on U.S. servers under legal frameworks markedly different from GDPR protections.

Fellou – ASI X

Billed as “the world’s first agentic browser,” Fellou offers operating system and file system access to manage end-to-end workflows. Given this scope, its privacy disclosures require careful examination.

Unfortunately, this information is relegated to a single footer link hosting the privacy policy.

The documentation reveals that Fellou reserves the right to access any data necessary to run the system: from geolocation to data entered on web pages, including local computer files when “Computer Use” permissions are granted:

All of this information is utilized to train the underlying LLM.

Furthermore, Fellou may share collected data with commercial partners to operate or improve services. Their policy explicitly notes:

“We will not collect sensitive information such as identification documents, bank accounts, or passwords. However, due to technical limitations and user interactions, we may unwittingly collect such data… Please do not input sensitive information into our services.”

This warrants serious deliberation before granting unrestricted access to enterprise or personal workstations.

ChatGPT Atlas – OpenAI

Atlas integrates ChatGPT Agent capabilities within a Chromium shell for macOS.

Granular privacy settings regarding memory retention and site-specific exclusion are featured prominently in the official documentation.

A notable feature is “disconnected mode,” which disables cookies and prevents account logins without explicit manual authorization.

Model training on user prompts is turned off by default unless globally enabled in the user’s primary ChatGPT account settings. While OpenAI’s historical approach to privacy has experienced shifts, Atlas provides accessible, granular controls.

3) Cyber Security and Autonomous Exploits

Finally, and crucially: cybersecurity.

Following the initial wave of LLM adoption, structural vulnerabilities quickly became apparent. Prompt injection was the first prominent vulnerability demonstrating how model instructions could be hijacked.

While partially mitigated, prompt injection remains an intrinsic byproduct of statistical probabilistic language architectures. The introduction of agentic browsers escalates these vulnerabilities exponentially.

In late October 2025, security firm LayerX identified an exploit allowing Atlas to execute arbitrary malicious payloads stored within ChatGPT’s contextual memory.

While developers work to implement safety boundaries (such as Meta’s “Agents Rule of Two” permission segmentation), the reverse vector poses equal danger: orchestrating fully autonomous cyberattacks.

In mid-November 2025, Anthropic confirmed detecting a large-scale cyberattack where Claude Code operated in agentic mode, autonomously executing up to 80% of steps typically requiring human hackers. Extending agentic browser versions to conduct autonomous reconnaissance and zero-day exploitation is an immediate risk.

Similarly, traditional CAPTCHA mechanisms are becoming obsolete. Evaluating human presence through warped text or image identification loses its purpose when multimodal browser models interpret visual context with near 100% accuracy.

What the Future Holds: The Dawn of the Agenticverse

While agentic browsers represent an avenue to monetize high compute costs, the true evolutionary milestone will occur when agentic systems interact cooperatively with one another.

Imagine a suite of AI agents operating on our behalf, communicating directly with counterpart agents representing other individuals, public administrations, and commercial enterprises.

We will inhabit an autonomous ecosystem executing operational and bureaucratic workflows at unprecedented speed without requiring constant manual oversight.

Consider this strategic shift: many draw parallels to the browser wars of the 1990s and 2000s (Netscape vs Internet Explorer; Chrome vs Firefox, Safari, and Edge). I see a fundamental distinction:

This new battle is not about controlling the software interface: it questions the very necessity of the World Wide Web as we have known it.

Today marks the three-year anniversary of ChatGPT’s public launch—the catalyst that transformed global AI awareness. Consider how much the landscape evolved in just 36 months.

Let’s take the inquiry further: will we still need web browsers?

What is the rationale for maintaining a traditional corporate website when the vast majority of stakeholder interactions are brokered by specialized AI agents? And if traditional web pages recede, what is the role of a browser?

We are entering the Agenticverse: an environment where transactions unfold in Agent-to-Agent mode, seamlessly executed on behalf of end users. We will instruct our agents to retrieve answers, book reservations, and negotiate contracts without ever launching a search engine.

This presents a future that is exceptionally powerful, demanding rigorous governance, ethical standards, and executive foresight.

Topics Covered
#Agentic AI#Agentic AI browser#Agenticverse
Ernesto Amato

Written by Ernesto Amato

Executive Leader and Fractional Chief AI Officer. Supporting Boards of Directors and C-Suite leaders in defining and implementing sustainable AI Governance, Enterprise AI Adoption, and Cultural Transformation.