Your Browser Just Got Dangerous. Boards Need to Pay Attention.

OpenAI (ChatGBT) dropped their new browser this morning. Within 2 hours, three executives texted me the same question: “Should we use this?” Wrong question. The right question: “Who owns decisions our browser makes with authenticated access to our systems?” I’ve been testing agentic browsers for 6 months-Dia, Perplexity’s Comet, now Atlas. Yesterday I told Dia: “Find cheaper flights for my Sydney trip and rebook if you save more than $200.” It did. Saved $340. Cancelled my original booking. Rebooked new flights. Then I realised, it just made a financial decision without showing me the new flight times. I’m now landing 4 hours later than my first meeting. The technology works. Brilliantly. But if I can’t trust it with a $340 flight decision after three months of testing, how ready is your organisation to give it access to your financial systems?

What Is an Agentic Browser?

Right, quick definition before we go further. Traditional browsers are passive. You Type. You search. You copy. You paste. You switch between 14 tabs to complete one task. Agentic browsers like Atlas and Comet collapse that process. You state intent. They research, synthesise, and act, all while authenticated as you. Three capabilities that matter: 1. Research Analyst Mode Open a competitor’s earnings report. Ask: “Summarise Q4 strategy in 100 words and compare to ours.” It reads both documents, extracts key differences, organises by Growth/Risk/Efficiency. No spreadsheet needed. It remembers context across sessions. Yesterday’s research informs today’s analysis without you re-explaining everything. 2. Strategic Advisor Mode After reviewing two company documents, ask: “Identify three strategic gaps between these plans.” It highlights where one relies on subscription revenue and the other on transactional volume. Drafts partnership opportunity bullets using sectors it remembers (mobility, infrastructure). 3. Autonomous Agent Mode This is where it gets dangerous, and powerful. You can tell it: “Find the executive contact page, confirm their email, create a calendar invite for 30 minutes next Tuesday.” It navigates, fills forms, pauses for your confirmation before sending. Or: “Draft a follow-up email in Gmail using the research context we just gathered.” It writes directly in the text box, you adjust tone (”more formal”), hit send. It has access to everything you have access to. The capability is high, The governance questions are harder. 30 Days Free - Comet Browser

The Four Concerns

Today’s agentic browsers solve problems. But they create four new governance gaps most boards haven’t considered: 1. The Credential Problem It authenticates as you. It books travel under your CFO’s login. Makes purchases with your procurement credentials. When it violates your travel policy or buys from unapproved vendors, who’s accountable? The CFO who deployed it? IT who configured it? OpenAI? 42% of Australian firms have no AI plans because governance uncertainty (Thomson Reuters 2025). Atlas just forced that decision forward by six months. 2. The Security Paradox Cambridge researchers just demonstrated prompt injection attacks on agentic browsers. Malicious code on a webpage can manipulate the agent, harvest credentials, exfiltrate data. Your browser can now authenticate as your executives and execute transactions autonomously. 3. The Memory Question It builds personalised memory over time. It remembers your preferences, your decision patterns, your organisational context. Sounds useful until you ask: Who audits what it remembers? How long is data retained? Where does Australian business data go during processing? 4. The Action Gap Traditional browsers wait for commands. Agentic browsers execute multi-step workflows autonomously. It’s documenting which decisions need human approval and which don’t.

Here’s My Advice

Don’t wait for perfect governance frameworks. Start learning now with tight constraints, BUT - Stop deploying blindly. **Stop Losing Control to Autonomous Actions: **Macquarie deployed AI tools to 3,000 employees. Timeline - 6 months. Not because the technology was hard. Because getting teams to trust AI-enforced standards while maintaining audit trails for ASIC compliance takes time.

Decision Matrix: Action Governance & Approval

Stop Ignoring Data Sovereignty: Atlas processes Australian business data through US infrastructure under OpenAI’s terms of service. For regulated industries, financial services, healthcare, legal - boards need to be aware. Before deploying to teams with access to sensitive data, answer these three questions:

• Where does our data go during processing?
• How long is it retained?
• What’s our ACCC position on cross-border data flows? For low-risk tasks (research synthesis, information gathering), deploy immediately. For high-risk workflows (financial transactions, contract modifications, client data), I’m waiting until we have clear data residency controls.

---