JP — ITmedia
Is It Risky to Do Other Tasks While Logged Into Online Banking? Seven AI Browsers, Including ChatGPT Atlas, Tested for Data Theft
Security2 min read
A research paper from the University of Washington tested seven "AI browsers" and similar tools. It found that four of them could let attackers steal private data, such as email and online banking information, from other sites open in the same browser.
What happened
Researchers at the University of Washington checked whether AI browsers are safe. These browsers have an AI agent that does tasks for you, like booking trips or reading web pages. The team found that four of the seven tools could be tricked into breaking a basic safety rule. This rule keeps one website from reading another website's data. An attacker could use this gap to steal sensitive information, such as bank details.
Key facts
- The team tested 7 AI browsers or browser add-ons.
- 4 of them had a path that let attackers get around the Same-Origin Policy.
- The Same-Origin Policy was introduced in 1995.
- ChatGPT Atlas (agent mode) was the first tool where the team successfully stole data in a test.
- ChatGPT Atlas stopped being offered in August.
Why the rule matters
Normally, a browser keeps each website in its own space. This is the Same-Origin Policy. Because of it, you can open a risky site in one tab and log in to your bank in another tab, and your bank data stays safe. An AI agent can move information between tabs for you. If it is given too much access, a hidden instruction on a bad website, called prompt injection, could trick it into sharing data from another site.
Why it matters
AI browsers are becoming more popular, and they can do many tasks for users. This study shows that their extra power can also create new risks. Even a careful user could have bank or email data stolen while using these tools. The study also warns that a fake bank phone number, once saved in the AI's memory, could later be given back to the user as if it were real.
What to watch
- Whether companies fix the gaps the researchers found.
- How Anthropic and Firefox respond, since they did not reply to the team.
- Whether Perplexity and OpenAI change their products, since they declined to receive the report.
- How safe AI browsers become before people trust them with bank and email tasks.
Glossary
- ChatGPT Atlas
- An AI browser made by OpenAI, which was tested in this study and stopped being offered in August.