Martin Jackson — your intelligent automation partner

Why I’m Not Touching AI Browsers (Well For Now)

Martin Jackson
Martin Jackson
Abstract pop art illustration featuring browser logos, a brain with circuits, and the text 'AI BrowserWars' in bold colours.

Everyone’s going mental about AI browsers right now. OpenAI’s ChatGPT Atlas, Perplexity’s Comet – they’re promising to revolutionize how we browse by having AI agents book your flights, fill out forms, and handle all your online stuff.

The tech press is calling it the future. I’m calling it bollocks .

Here’s What 30 Years Taught Me About Shiny New Tech

Look, I’ve been building AI systems since 2021 – back when it was proper mental and nobody knew what they were doing. My first system was supposed to grab form data, transcribe details, update spreadsheets, and push everything into our CRM.

Did it work perfectly? F*ck no.

It would run like a dream for weeks, then mysteriously decide that “email address” meant “favourite pizza topping” and we’d spend hours figuring out why our CRM was having a complete meltdown.

But here’s the thing – when my early AI system failed, it created data entry errors . Annoying? Yes. Business-ending? No.

When AI browsers fail? They can empty your bank account while you sleep .

The Security Nightmare Nobody’s Talking About

Collage-style illustration of glass shards and lines in yellow, red, blue, and black, with the text 'Infected Signal' visible.

While everyone’s getting excited about AI doing their shopping, security researchers are discovering some terrifying vulnerabilities:

Prompt Injection Attacks : Malicious websites can embed hidden commands that trick AI browsers into doing stuff you never asked for. Visit the wrong site and boom – your AI’s copying cookies, clicking dodgy links, or transferring money without you knowing.

Screenshot Hijacking : Researchers found that attackers can hide commands in images. When your AI browser takes a screenshot, these hidden instructions execute like they’re legitimate commands. Mental, right?

The Memory Problem : LayerX Security discovered that attackers can inject malicious instructions into ChatGPT Atlas’s memory system. These “tainted memories” stick around forever and can make your AI execute harmful code whenever you use it.

The Search Bar Trap : Atlas combines search and commands in one bar. Crafty attackers can create special links that bypass safety checks, making injected instructions run with full trust. It’s like giving a stranger the keys to your house.

Why This Feels Like Déjà Vu All Over Again

I’ve watched this movie before, dude. Thirty bloody years of businesses rushing into new tech without understanding the risks:

  • Email without understanding phishing
  • Websites without understanding security
  • Social media without understanding reputation management
  • Cloud computing without understanding data protection

Every time, the early adopters got burned . Hard.

The difference this time? The stakes are way higher . We’re not talking about embarrassing email mistakes – we’re talking about AI agents with access to your banking, your customer data, your entire business system.

The Business Risk Nobody’s Calculating

Retro geometric illustration of an orange, blue, and red starburst explosion with angular shapes and the text 'BOA / do.m'.

Current AI browsers need massive access to work – emails, calendars, contact lists, login credentials. They’re basically asking for the keys to your entire digital life.

Security experts are warning that normal web safety measures don’t apply here. If an AI agent can pretend to be you when you’re logged into sensitive accounts, it could:

  • Make unauthorized transactions
  • Access confidential customer data
  • Compromise your entire business system
  • Give attackers a permanent backdoor

Brave’s research team put it perfectly: “Simply summarizing a specially crafted Reddit post could result in an attacker being able to steal money or your private data.”

That’s not a feature – that’s a bloody nightmare .

Even the Companies Building This Stuff Know It’s Broken

OpenAI had to create a “logged out mode” to limit data exposure. Perplexity built real-time detection systems for attacks. But security professionals are clear: these browsers aren’t foolproof .

The core problem? AI systems can’t reliably tell the difference between your legitimate instructions and malicious commands hidden in web content. This isn’t a bug they can fix – it’s a fundamental limitation of how these systems work right now.

What Smart SMEs Should Do Instead

Geometric, colourful illustration of a sphere breaking into bars and fragments with red, blue, yellow, and black textures. Text reads: 'Pragmatic Pathways | Pop Art Collage Style | flux-kontext-max'.

  1. Stick to Proven Stuff : Use established automation tools like MAKE, Zapier, or N8N. They’ve got mature security and clear boundaries between your input and external data.
  2. Start Small with AI : Try contained AI applications – customer service chatbots, content generation, data analysis. Stuff that doesn’t need access to your entire digital life.
  3. Ask the Hard Questions : Before adopting any AI tool, ask: What data does it access? How does it handle logins? What happens if it gets compromised? Can we limit what it can do?
  4. Let Others Be the Guinea Pigs : Wait for security to mature. Let the big companies figure out the problems and pay for the fixes.

The Bottom Line

AI browsers are interesting tech , but they’re not ready for business use. The security risks massively outweigh any convenience benefits, especially for SMEs that can’t afford a proper security breach.

I’ve seen this pattern for 30 years : promising technology gets hyped before it’s ready, early adopters pay the price, and smart businesses wait for the dust to settle.

The businesses that thrive aren’t the ones adopting every shiny new thing first – they’re the ones adopting the right technology at the right time , with proper safeguards in place.

Don’t be the guinea pig with your business bank account.

Ready to explore AI automation that actually works safely? Let’s chat about proven alternatives that deliver results without the security nightmares.

Tags: , , , ,

Latest Posts:

AI Disclaimer: Because Apparently We Need One Now

Right, let’s get this out of the way before someone has a proper meltdown…

Yes, I use AI. Shocking, I know. An AI automation specialist using AI – it’s almost like a chef using knives or a plumber using spanners. Mental, isn’t it?

Here’s the thing – I’ve never pretended to be Picasso. I can barely draw a decent stick figure, and my artistic education peaked at colouring inside the lines (which I was rubbish at). So when it comes to creating visuals or researching content, AI is bloody brilliant at doing the heavy lifting.

But here’s what AI can’t do – it can’t replicate 30+ years of getting my hands dirty in real businesses. It can’t copy the scars from watching “simple integrations” turn into six-month nightmares. And it definitely can’t swear like a proper Brit who’s spent decades dealing with systems that mysteriously decide important data fields mean something completely different.

The EU’s Having a Field Day (Naturally)

Speaking of overreactions, the EU’s now mandating AI content labels because apparently they’d regulate the opening of a crisp packet if they could. Their new AI Act requires everything to be marked, watermarked, and probably stamped in triplicate.

Meanwhile, the UK’s taking a more sensible approach – “use your judgement, don’t be a pillock about it.” Refreshing, really.

But here’s the kicker: research shows that slapping “AI MADE THIS” labels on everything actually makes people trust it less – even when the content is identical quality. It’s like putting a “CONTAINS CHEMICALS” sticker on water. Technically true, completely mental.

Remember When Social Media Was “Just a Fad”?

This whole AI panic reminds me of those highly respected business experts who dismissed social media as a passing trend back in 2006. You know the ones – sitting in boardrooms saying “Facebook is just for university kids” while their junior staff begged them to pay attention.

Those same companies spent years trying to catch up after realizing they’d missed the boat entirely. By the time they finally joined Facebook in 2011, their competitors had built communities, refined strategies, and owned the conversation.

Here’s the scary bit: AI adoption is happening three times faster than social media did. We’re already seeing the same pattern – 5% of companies are crushing it with AI while 95% are still “evaluating options” and “waiting to see how it develops.”

Guess which group my clients want to be in?

My Process is Simple:

  1. AI helps me research and draft the boring bits.
  2. I add the personality, experience, and occasional profanity.
  3. I make sure it actually makes sense for real businesses.
  4. I take full responsibility for everything published.

The Reality Check

To those clutching their pearls about AI assistance: get over it. We’re not going back to typewriters and encyclopaedias. The future belongs to people who embrace tools that make them more effective, not those hiding behind “authentic struggle.”

Besides, if you think AI came up with my stories about early automation systems going mental and mixing up data fields… well, you’ve clearly never worked with AI long enough to know it’s nowhere near that creative (or sweary).

TLDR: AI helps, humans create, everyone wins. Now can we get back to actually helping businesses instead of arguing about tools?

P.S. – If you’re still bothered by this, you’re probably not my target audience anyway. My clients care about results, not whether I used a calculator or did the maths by hand. And they definitely don’t need the EU telling them how to label their content like it’s a tin of beans.

P.P.S. – Those “experts” who missed social media? They’re probably the same ones telling you to “wait and see” with AI. How’d that work out last time?

All insights are based on publicly available information and my 30+ years of business technology experience, including building AI systems since 2021. This content is for informational purposes only and is not intended to be professional or legal advice.

More from the Journal