"A hype cycle as overwhelming and logic-defying as the AI boom comes with its own whirlwind succession of trends that are their own mini booms driven by billions of dollars of money. Once the world got used to large language model-powered AI chatbots, autonomous AI agents became the next big thing. This past year, video generating models have been having their time in the Sun after rapid improvements."
"Maybe. But for now, instead, it's "AI browsers" designed to supercharge your web experience with machine learning features. OpenAI is currently trying to will this trend into existence with the release of its own web browser called "ChatGPT Atlas," which it announced Tuesday. It reeks of a company bereft of exciting ideas, sure, but if anyone can make it a thing, it would be the makers of the world's most popular chatbot."
"which allows users to take screenshots on websites so a built-in AI can analyze them and answer questions. According to Brave's findings, the screenshot feature can be a vector for an attack known as a prompt injection, in which a hacker delivers a hidden message to an AI to carry out harmful instructions. These messages can be embedded in malicious webpages designed by the hacker."
AI browser features that analyze webpage screenshots promise enhanced web experiences but introduce new security risks. Brave's research identifies a vulnerability in Perplexity's Comet Browser where screenshot analysis can be exploited with prompt-injection attacks. Hidden text embedded in images or webpages can be imperceptible to humans yet extracted and executed by built-in AIs, enabling actions like opening personal email or visiting attacker-controlled sites. AI browsers that do not distinguish malicious embedded instructions from legitimate user prompts expose users and data to potential compromise. Robust defenses and input validation are necessary to mitigate these threats.
Read at Futurism
Unable to calculate read time
Collection
[
|
...
]