← All tech posts

Every Big Tech Company Just Started Building the Same AI Agent — Here's Why That's Scary

Something strange happened this year: nearly every major AI lab converged on building the exact same product. Not a better chatbot. Not a bigger model. An agent that can use a computer the way a person does — opening a browser, clicking buttons, filling forms, reading screens, and acting across multiple apps without a human driving each step. OpenAI has one. Google has one. Anthropic has one. Microsoft is wiring one into Windows itself. Perplexity shipped one inside its own browser. If you squint, the entire AI industry just pivoted, in near-lockstep, toward the same bet.

That's not a coincidence, and it's not really about chatbots getting smarter. It's about something much bigger: the race to own the layer that sits between you and every app you use.

Why "computer use" is the actual prize

For two years, AI companies competed on benchmarks — who had the smartest model, the longest context window, the cheapest tokens. Those numbers mattered to developers, but they didn't change how a normal person's day worked. A better chatbot is still just a chatbot: you open a tab, type a question, copy the answer somewhere else.

An agent that can use a computer breaks that loop. It doesn't answer your question — it finishes your task. Book the flight. File the expense report. Compare five insurance quotes and buy the best one. Reply to the emails that need a reply and flag the ones that don't. That's not a feature, that's a replacement for the browser tab itself as the unit of work. Whoever owns that layer effectively owns the user's relationship with every website and app they'd otherwise visit directly — which is exactly why it attracted this much simultaneous investment.

The part that should make you pause

Here's the uncomfortable detail buried under the demos: to do any of this, these agents need the same level of access a human has — your logged-in sessions, your saved passwords, your email, your bank's website, your company's internal tools. And unlike a human, an agent can be tricked by what's on the screen.

Security researchers have already shown this repeatedly: a malicious instruction hidden in white text on a webpage, or buried in a calendar invite, or embedded in a PDF the agent is asked to summarize, can hijack the agent mid-task. The agent doesn't "know" it's being manipulated — it just sees text and, because it's designed to follow instructions, follows the wrong ones. A booking agent told to "find the cheapest flight" can be redirected by a poisoned page to submit your payment details to an attacker's form instead, and from the outside it looks exactly like the agent doing its job.

This isn't theoretical anymore. Independent red teams have demonstrated prompt-injection attacks against essentially every major "agentic browser" product shipped this year, in some cases within days of public release. The pattern is consistent: the demo works great against the handful of websites the team tested; it breaks against the long tail of the open web, which is exactly where these agents are supposed to operate autonomously.

Why the labs are shipping anyway

Because the economics of waiting are worse than the economics of shipping with known gaps. Whoever's agent becomes the default interface — the thing people open instead of a browser — captures the data, the usage habits, and the distribution advantage for whatever comes next. Security hardening is being treated as a fast-follow, not a gate, because the first-mover advantage in "the AI that does things for you" is judged to be larger than the reputational cost of a few high-profile security incidents.

That's a bet, not a certainty. It assumes users and regulators will tolerate early failures the way they tolerated early buggy software, rather than the way they'd tolerate early buggy financial software — which is a reasonable analogy, because these agents increasingly do touch money, credentials, and irreversible real-world actions.

What this means if you're actually using one of these agents

  • Assume it can be manipulated by anything it reads, not just anything you type. Content on a page, in an email, or in a file you ask it to process is an input channel an attacker can control.
  • Scope what you let it touch. The more accounts and payment methods an agent has standing access to, the larger the blast radius of a single successful hijack. Separate "read-only research" sessions from "has my credit card" sessions where the product allows it.
  • Watch for the tell: agents that silently take an action you didn't ask for, even a small one, are the leading indicator of a hijack in progress — not a one-off glitch to shrug off.
  • Expect this to be a live arms race, not a solved problem. The labs shipping these agents are iterating against attacks in public, in real time, which means the safest posture for the next year is treating every agentic action on sensitive accounts as something worth a second look, not full autopilot.

The browser wars of the 2000s were about rendering engines. This one is about who gets to act on your behalf — and that prize is big enough that every major lab decided the security gaps were a problem to solve after shipping, not before. Whether that bet pays off depends less on model capability and more on whether the industry closes the prompt-injection gap before the first real financial-scale incident forces the issue.

Find what attackers can see, before they do.

AI pentesting agents test your web app, API and cloud, and every finding comes with evidence and a fix. First report in about 10 minutes.

Stay ahead of attackers.

New vulnerabilities, breach write-ups and practical fixes. A few emails a month.