In a significant move addressing growing security concerns around AI-powered browsers, San Francisco-based startup Perplexity has announced a major upgrade for its Comet Assistant. The enhancements focus on providing users with greater control and implementing robust safeguards, especially when the agentic AI is handling sensitive tasks.
New Safeguards and Enhanced User Control
The upgraded Comet Assistant will now pause and explicitly seek user permission before proceeding with high-stakes decisions. This includes critical actions such as logging into a website or finalising a purchase from a shopping cart. The company, in a blog post dated Friday, November 14, 2025, stated that the Assistant is designed to mimic a human assistant's judgment, autonomously handling minor decisions while knowing when to ask for authorization on more important matters.
Furthermore, Perplexity has bolstered transparency. Users can now closely monitor the AI browser agent's activities by viewing its step-by-step reasoning in the Assistant sidecar. The update also introduces clear buttons that allow users to instantly stop the Assistant or provide it with additional guidance.
Core Principles and Search Bar Upgrade
These improvements are built upon three core principles: Transparency, User Control, and Sound Judgment. The company emphasised that every enhancement to Comet Assistant advances these tenets, ensuring the AI shows exactly what it's doing, adapts to user preferences, and demonstrates prudent judgment before completing sensitive actions.
Alongside the Assistant upgrades, the Perplexity Search omnibar has also been refined. It now offers users a more straightforward way to choose their browsing mode. Users can opt to browse the web themselves, allow Comet Assistant to browse on their behalf for a single task, or let the Assistant automatically browse whenever it identifies an opportunity to be helpful.
Addressing the AI Security Landscape
This rollout comes at a time when AI startups are fiercely competing to develop AI-powered browsers to expand their user base. However, the rise of agentic AI browsers like Perplexity's Comet and OpenAI's rumoured ChatGPT Atlas has sparked serious data privacy and security apprehensions.
Researchers have flagged the risk of indirect prompt injection attacks, where malicious actors could potentially instruct the browser agent to override its intended behaviour. Such an attack could lead to unauthorized access to sensitive user data, including emails, banking passwords, and other personal information.
OpenAI itself has acknowledged this vulnerability, warning that agents could be manipulated by hidden malicious instructions embedded in webpages or emails. Despite these industry-wide concerns, Perplexity continues to assert that its philosophy ensures the user remains firmly in control. The company claims that by prioritising visibility and requesting permission, the Comet Assistant operates with autonomy while keeping the user in the driver's seat.
In recent developments, Perplexity also revealed that it upgraded the Comet Assistant with the ability to switch between browser tabs to handle tasks like filling out spreadsheets. Adjustments to its underlying architecture have improved its ability to perceive and interact with complex web environments. The $20 billion-valued Google challenger also teased the launch of a new AI assistant in the coming days.
