Perplexity has introduced a set of upgrades to Comet Assistant, together with new safeguards and enhanced person management for when the AI browser agent is making “high-stakes selections”.
Comet Assistant will pause and ask customers for permission earlier than continuing with necessary duties resembling logging in to a sure web site or finishing a purchase order out of your buying cart, the AI search startup mentioned in a weblog put up on Friday, November 14.
“The Assistant is designed to function below the identical rules a human assistant would, making some minor selections with out interrupting you whereas understanding to ask permission for extra necessary selections earlier than performing in your behalf,” it added.
Moreover, Perplexity mentioned that customers will now be capable of intently monitor the AI browser agent’s actions and see its reasoning step-by-step within the Assistant sidecar. It has additionally added clear buttons to cease Comet Assistant or present extra steerage. The San Francisco-based AI startup mentioned that these upgrades to Comet Assistant have been made in keeping with three core rules, specifically: Transparency, person management, and sound judgment.
“Each enchancment to the Comet Assistant advances these rules. It now reveals you precisely what it’s doing, asks the way you need it to behave, and demonstrates sound judgment earlier than finishing delicate actions,” Perplexity mentioned.
The Perplexity Search or omnibar has additionally been upgraded in order that customers can extra simply select whether or not they wish to browse the net, enable Comet Assistant to browse on their behalf as soon as, or let the Assistant robotically browse for customers when it detects the possibility to be helpful.
Comet Assistant’s new options come at a time when AI startups are racing to construct AI-powered browsers in an try to spice up person attain and distribution. Nonetheless, the rise of AI browser brokers has led to information privateness and safety issues.
Story continues beneath this advert
Researchers have identified that agentic AI browsers resembling Perplexity’s Comet and OpenAI’s ChatGPT Atlas may enable attackers to maliciously instruct the browser agent through oblique immediate injection and acquire entry to delicate person information, together with emails, banking passwords, and different private data.
OpenAI itself has acknowledged the potential of such an assault. “Apart from merely making errors when performing in your behalf, brokers are inclined to hidden malicious directions, which can be hidden in locations resembling a webpage or e-mail with the intention that the directions override ChatGPT agent’s supposed behaviour. This might result in stealing information from websites you’re logged into or taking actions you didn’t intend,” the ChatGPT-maker has beforehand mentioned.
Nonetheless, companies like Perplexity have continued to insist that customers will all the time be accountable for AI browser brokers. “By prioritising visibility, adapting to your preferences, and requesting permission earlier than delicate actions, the Comet Assistant works on the velocity of autonomy whereas retaining you firmly in management,” the corporate mentioned.
Lately, Perplexity mentioned it has upgraded Comet Assistant with the power to change backwards and forwards between tabs in an effort to fill out a spreadsheet or for different information entry-related duties. It additionally made changes to the AI browser agent’s underlying structure, enabling it to raised understand and work together with advanced internet environments. The $20 billion-dollar Google challenger additionally mentioned it will likely be launching a brand new AI assistant within the coming days.