AI Predictions for 2026
Happy new year! The purpose of this post is to write out ideas, thoughts and predictions regarding AI trends for 2026.
My main prediction is that Agentic AI will become a mainstream tool for everyday users and there will likely be some societal growing pains (e.g. security). By Agentic AI here I mean AI that can perform actions ("do things") as opposed to simply describe things. There is an important conceptual difference between passively observing/describing the world versus actively imparting change upon the world. Currently, generative AI is utilized by most people as a browser tool or phone-based app where LLMs are engaged to produce natural-language output (ChatGPT) or in some cases multimodal visual output like images or videos (Sora). Ultimately, utilizing this output to execute actions in the world is left to the human user.
For example, if you are planning a trip to Mexico, you might engage your favorite LLM in the browser to ask for recommendations on where to stay and what to do, perhaps even request an itinerary be written for you. However, it is still on you as the human user to actually book reservations at hotels, flight tickets and so forth. The AI experience then falls short of a true travel agent who can execute booking the trip for you. In 2026, I predict AI Agents will reach a state of mainstream adoption and maturity where they can actually book your trips for you and execute tasks with real-world impact, serving as digital travel agents or digital butlers rather than simply information generators.
Perplexity's Comet browser stands out as an example of an Agentic AI native application that can perform various tasks for users:
- organize a meeting
- write an email based on the opened website
- buy something
- list today's tasks
However, in order to complete the above, users must give Comet access to many aspects of their personal information including: calendars, email inboxes, credit card information etc. This introduces security vulnerabilities if hackers are able to engage the underlying AI Agents (prompt injection) to steal sensitive personal information or perform compromising tasks such as transfer funds.
According to Copilot, the Chrome and Edge browsers are not currently set up to execute tasks for users via Agentic AI unlike Comet. In the cases of Chrome and Edge, AI's role is relegated to the sidebar as a chat assistant which can process information of interest to the user but does not execute material tasks in the browser. However, Anthropic is working on a Chrome extension to embed its Claude model into the browser as an Agentic AI assistant. Anthropic's internal security testing of the extension revealed a 23.6% security risk for the tool to be abused in the absence of guardrails. Anthropic is continuing to work on mitigating the associated security risks before providing its browser extension for mainstream use (for now it is a "research preview").
References:
Perplexity's own browser "Comet" with integrated AI agents now available | heise online
Perplexity's AI Comet Browser Is Now Available for Free for Everyone | Lifehacker
Your AI Browser May Be Vulnerable to 'Prompt Injection' Attacks | Lifehacker
Anthropic’s auto-clicking AI Chrome extension raises browser-hijacking concerns - Ars Technica
Anthropic launches a Claude AI agent that lives in Chrome | TechCrunch
Comments
Post a Comment