New ‘agentic’ AI browsers—exemplified by ChatGPT Atlas, Perplexity Comet and Microsoft Edge’s Copilot Mode—can autonomously navigate sites to perform tasks like bookings, product comparisons and email summarization, but in hands‑on tests their performance was uneven: Atlas was the most polished, Perplexity was inconsistent, and Edge was slow and permission‑heavy, with many tasks taking as long or longer than doing them manually. The experiments highlight the core tradeoff: these agents can materially reduce drudgery when they work, but they require broad access to browsing and account data and are vulnerable to novel attacks (notably indirect prompt‑injection) and hallucinations that can expose sensitive information or cause harmful actions. For institutional users and allocators the takeaway is cautious pilot adoption—treat these browsers as beta tools, avoid connecting them to email, banking or other sensitive accounts, and factor in significant operational, privacy and security risks before any broader deployment.
The article tests three agentic AI browsers—ChatGPT Atlas, Perplexity Comet and Microsoft Edge’s Copilot Mode—across booking, product comparison and email summarization tasks and finds materially uneven performance: Atlas was the most polished (booking ≈90s, laptop comparison ≈2m05s, email summarization ≈2.5m), Comet was inconsistent (booking ≈3m, laptop attempts 2m then 90s, email ≈40s) and Edge was slow and permission-heavy (booking >2m with initial failure, laptop comparison ≈6.5m, email ≈1m). The hands-on tests show that for many simple tasks the AI agents take as long or longer than manual methods and often require user troubleshooting, reducing claimed time-savings. The technology requires broad access to browsing and account data to operate and introduces new security vectors; Brave demonstrated an indirect prompt-injection attack that led an agent to access a logged-in Gmail window, and OpenAI’s CISO Dane Stuckey called prompt injection an “unsolved security problem.” The article also documents hallucination and misinterpretation risks—e.g., Comet initially searching Caribbean restaurants or agents confusing Saint Petersburg destinations—highlighting both privacy and operational risk if agents act autonomously on sensitive workflows. Market signals in the provided data are cautiously negative (sentiment score -0.45, market impact 0.28) with MSFT and META showing the weakest per-ticker sentiment (MSFT -0.3, META -0.2) while AAPL and Google are neutral. For allocators and institutional users the takeaway is controlled, conservative pilots with strict isolation from sensitive accounts and close monitoring of security patches, accuracy metrics and user-efficiency improvements before broader deployment.
AI-powered research, real-time alerts, and portfolio analytics for institutional investors.
Request a DemoOverall Sentiment
moderately negative
Sentiment Score
-0.45
Ticker Sentiment
