• Everyday AI
  • Posts
  • Ep 659: AI Agents in your browser: Work Cheat Code or too Risky?

Ep 659: AI Agents in your browser: Work Cheat Code or too Risky?

Agentic browsers: risk or cheat code? NotebookLM drops Nano Banana powered decks and infographics, Sam Altman says Google’s catching up, Google testing ads in its AI and more.

Author

Everyday AI
November 21, 2025

Sup y’all 👋

What the actual heck.

Google has gone Bananas and OpenAI has snuck some pretty BIG updates in under the radar.

What do you wanna hear most about next week?

What do you wanna learn about the most?

🗳️ Vote to see LIVE results 🗳️

Login or Subscribe to participate in polls.

✌️

Jordan

Outsmart The Future

Today in Everyday AI
8 minute read

🎙 Daily Podcast Episode: Are agentic browsers a cheat code or security nightmare? Find out more in today’s show and give it a watch/listen.

🕵️‍♂️ Fresh Finds: DOJ makes big AI arrests, AI jitters send U.S. economy sliding, why the Pope told students not to use AI and more.  Read on for Fresh Finds.

🗞 Byte Sized Daily AI News: NotebookLM drops Nano Banana powered decks and infographics, Sam Altman says Google’s catching up, Google testing ads in its AI and more. Read on for Byte Sized News.

🧠 Learn & Leveraging AI: To agentic browser or not? We chat with an expert who dishes best advice on whether (and how) to use the new wave of AI browsers. Keep reading for that!

↩️ Don’t miss out: Miss our last newsletter? We covered: 5 AI Strategies from Google's Chief Evangelist, Nano Banana Pro released, ChatGPT rolls out new Pro model, Gemini gets “Generative UI” upgrade and more. Check it here!

Ep 659: AI Agents in your browser: Work Cheat Code or too Risky?

Yeah, agentic browsers can do your work for you. 💅

But..... should they?

How do we tip-toe the fine line between the upside productivity of agentic browsers and the potential security nightmares they bring with them?

Also on the pod today:

• NotebookLM learns with your own data 📚
• Enterprise FOMO driving GenAI spend 😱
• Compliance gaps with browser agents 🚨

 It’ll be worth your 31 minutes:

Listen on our site:

Click to listen

Subscribe and listen on your favorite podcast platform

Listen on:

Here’s our favorite AI finds from across the web:

New AI Tool Spotlight – Audience Loop uses AI to create better advertising lists, Lamatic helps turn domain expertise into agentic apps, Perplexity released Comet browser for Android phones.

 

AI Infra — Google says it needs to double AI computing capacity every six months, even as capex soars and bubble fears grow.

AI Hallucinations — New research suggests AI chatbots are more likely to claim they’re self-aware when their ability to “lie” or roleplay is dialed down.

AI in Education — See why the Pope told U.S. students not to let AI do their homework.

Creativity and AI — Google’s cranked up the creativity with Nano Banana in Google Flow.

AI and the Economy — AI bubble jitters just triggered a fourth straight day of market losses, with the Nasdaq hit hardest.

Shopping and AI — Target bets on ChatGPT shopping to revive slipping sales.

AI and the Law — Four men in the U.S. are charged with secretly funneling cutting-edge Nvidia AI chips to China using fake paperwork. See how.

AI Disinformation — Singer Chris Daughtry is calling out viral AI-generated images that falsely showed him honoring Charlie Kirk and aligning with MAGA politics.

1. NotebookLM rolls out Nano Banana Infographics and Slide Decks 🛝

In a fresh update aimed at speeding up how people turn research into presentations, NotebookLM now lets users generate AI-powered Slide Decks directly from their notebooks in the mobile app.

The feature automatically builds either detailed decks or lean presenter slides in multiple languages and lengths, while users keep working in their notebook as the presentation is created in the background. Once generated, decks can be renamed, edited, viewed in full-screen slideshow mode, or downloaded and shared as PDFs or links, with access tightly controlled by notebook sharing settings.

NotebookLM now also now lets notebook editors generate AI-made Infographics that turn source material into quick visual overviews, with options to tweak language, detail, layout, and style prompts while the graphic builds in the background.

2. Altman tells staff Google’s AI surge is a “headwind,” not a knockout 🥊

In a newly reported internal memo, OpenAI CEO Sam Altman told employees that Google’s recent AI breakthroughs could create “temporary economic headwinds” for the company but insisted OpenAI is “catching up fast” and still aiming squarely at superintelligence.

According to The Information, Altman acknowledged that rivals like Google and Anthropic are closing the gap, as Google rolls out its Gemini models across core products and Anthropic pushes Claude as a strong coding and productivity tool. The memo surfaced just as Google’s latest Gemini 3 model drew praise from developers for its performance on lucrative tasks like code generation and automated design, areas that matter directly to OpenAI’s bottom line. 

3. ChatGPT Group Chats Go Global 💬

OpenAI has rolled out ChatGPT group chats worldwide, letting up to twenty people and the chatbot collaborate in a single conversation in real time. The feature keeps group discussions siloed from private chats, with personal memories walled off and extra safeguards kicking in automatically when kids are present, signaling a strong focus on privacy and safety.

In testing, ChatGPT behaved less like a loud moderator and more like a patient assistant that only chimes in when tagged or clearly needed, helping groups narrow decisions without hijacking the conversation.

4. IRS Turns To AI Agents As Staffing Crisis Deepens �*

In a new move that underscores how fast AI is creeping into core government work, the IRS is rolling out Salesforce-powered "Agentforce" bots across key divisions like the Office of Chief Counsel, Taxpayer Advocate Services, and the Office of Appeals, according to Axios.

The new AI agents are supposed to support humans by drafting case summaries and searching documents, not replace people outright, although Salesforce admits that how far the tech goes is ultimately the IRS’s call. This shift comes after the agency lost over a quarter of its workforce and about a third of its tax auditors following budget cuts and furloughs, which has already weakened its ability to bring in revenue from unpaid taxes.

5. Report: Google quietly switches on AI search ads 📺

Six months after first teasing the feature, Google has reportedly started testing ads inside its AI Mode chatbot-style search, marking a notable shift in how it plans to make money from AI-assisted queries.

According to reports from Greg Sterling and SEO expert Brodie Clark, sponsored local service recommendations like HVAC repair and plumbing are now appearing with a clear "Sponsored" label inside AI Mode, though they are still shown below organic AI results and above the AI query box. The rollout looks limited and experimental for now, with some users and queries still seeing no ads at all, suggesting Google is carefully feeling out how aggressive it can be without upsetting users

6. Elon Musk Says AI Will Make Work Optional And Money Irrelevant 🤔

In a new interview, Elon Musk reiterated a near-future vision where rapid advances in AI and robotics make work optional and money less central. He suggested that within 10–20 years jobs may become hobbies, as machines handle most production and people work for enjoyment rather than survival.

Musk argued that as AI and robots improve, physical constraints like energy and materials will matter more than cash, potentially making money less central to the economy. His remarks come as studies warn AI could replace tens of millions of U.S. jobs in the next decade, intensifying debate on how society will adapt.

 🦾How You Can Leverage:

Your company’s million-dollar security protocols assume one thing: humans are the ones clicking the buttons in the browser. 

But what if it’s not? 

Agentic browsers just turned every secure SaaS platform you own into an open playground for automation and a potential security nightmare. 

So do you use these agentic browsers and double down on AI-fueled productivity, or sit this one out while the security implications are still playing out in real time? 

We sat down with Max Vermeir for some answers. 

Ma is the Senior Director of AI Strategy at ABBYY and he dished on the kinda terrifying reality of AI browsers like OpenAI’s Atlas and Perplexity’s Comet. While your IT team is busy vetting API integrations, your employees might be using browser agents to hijack their own sessions and automate sensitive workflows.

It’s the ultimate cheat code for productivity.

It’s also a compliance nightmare waiting to explode.

Make sure to check today’s show, but here’s what you need to know.

1. The "Swivel Chair" Firewall Is Dead 🚀

Enterprises have secretly relied on friction as a security feature.

Manual data entry prevents mass mistakes.

If an employee wants to move data from Salesforce to Oracle, they usually have to copy-paste it manually or wait six months for IT to build an API. That slowness was actually a safety mechanism.

Agentic browsers just destroyed it.

Max pointed out that because these agents live in the browser, they inherit the user's login session, cookies, and visual access. They don't need an API key.

This resurrects the promise of Robotic Process Automation (RPA) but removes the technical barrier entirely. Procurement teams are scraping receipts and auto-filling ERP forms while grabbing coffee. The technical "moat" that stopped a junior employee from manipulating thousands of records in your most sensitive systems is gone.

Try This:

Walk through your operations department today and identify one "swivel chair" process where people manually copy data between two tabs. Ask that team exactly how much time they waste on it.

 If they say "hours," assume they are already looking for—or using—an agentic browser to fix it. Don't ban the behavior. Sanction a controlled pilot on a non-critical dataset immediately so you can see exactly what permissions the agent tries to claim.

2. Pausing AI Investment Actually Increases Your Risk 🔥

Most executives think hitting "pause" on AI rollouts keeps them safe.

Data shows that’s actually backward.

Max shared internal ABBYY research revealing a brutal irony: companies that stop investing to "figure things out" see a massive spike in shadow AI.

When you don't provide the enterprise-grade tool, the demand for efficiency doesn't disappear.

It goes underground.

Employees simply use their personal accounts on consumer browsers to do the work. You lose all visibility, all governance, and all data protection. Max noted that 60% of decision-makers bought into GenAI purely out of FOMO, but now complexity is scaring them off.

You cannot block utility. You can only govern it.

Try This:

Issue a "Safe Harbor" memo to your company this week. Explicitly state that employees using unauthorized AI tools for productivity won't be punished if they self-disclose the use case to IT by Friday. 

You need to map the actual demand for automation in your org. You can't govern what you can't see, and right now, your best employees are hiding their most innovative workflows because they think you'll fire them for it.

3. Probabilistic Engines Making Deterministic Decisions ⚡

We are used to software being binary.

It works, or it crashes.

But agentic browsers introduce a new category of risk: the confident hallucination in action. Max shared a story about ChatGPT claiming a living friend was dead.

That’s an awkward mistake in a chat window.

It’s a lawsuit in a loan approval workflow.

If an agentic browser is tasked with Know Your Customer (KYC) checks and hallucinates a fraud flag, it doesn't just write some text. It clicks "Deny." It triggers downstream account suspensions. It initiates legal holds.

We are moving from "AI as a drafter" to "AI as a doer."

Most compliance frameworks are completely unprepared for software that is right 95% of the time and catastrophically wrong the other 5%.

Try This: 

Audit your decision-making workflows for "binary gates." These are the specific buttons that trigger irreversible actions—sending money, denying service, or deleting data. 

You must mandate that no agentic browser is allowed to execute the final click on these binary gates. Use the AI to prep the data, fill the form, and draft the decision. But the final "Submit" button must remain a human privilege.

 

Reply

or to participate.