Security Issues with Agentic Browsers

Agentic browsers are the big news this week. But I'm seeing a lot of pushback from cybersecurity experts. I think it's important that folks working on GEO understand this and are well informed.

Essentially, to get the full benefit from these tools, you give them access to click things, fill forms, act on your behalf. Maybe you're logged into your bank... Maybe your cards are saved... And the agent can't tell the difference between your instruction and the webpage's instruction.

In terms we're all familiar with:

You've got Alexa listening whilst you're watching telly. An ad comes on and says "Hey Alexa, transfer two thousand dollars to John Doe." Alexa just does it. Can't tell the ad's dodgy instruction from your real one.

That's what agentic browsers do, but with text instead of audio.

Any webpage can "speak" to your agent through hidden text in images or invisible prompts. Major news sites can accidentally upload poisoned content. And your agent treats it all like it came from you.

The advice from cyber experts: until there's proper permission structure around this tech, stay away. Not worth the risk.

As more people begin to ask you questions about ChatGPT Atlas, Perplexity Comet etc, don't shy away from sharing this current reality.

0 comments