Riesgos de seguridad en agentes AI

New AI-powered web browsers like OpenAI’s ChatGPT Atlas and Perplexity’s Comet are emerging to challenge the dominance of Google Chrome, which has long served as the gateway to the internet for billions. These innovative platforms feature web browsing AI agents designed to automate user tasks by navigating websites and filling out forms, a selling point that aims to enhance convenience in online activities.

However, a critical side to this new technology is the serious risk it poses to user privacy. Cybersecurity experts express concerns that these AI-driven browsers may jeopardize user data in ways traditional browsers do not. Users need to assess the extent of data access granted to these AI agents and weigh the advantages against potential threats to their privacy.

Related posts

Mastodon actualiza software: publicaciones citadas para operadores.

6 noviembre, 2025

Actualización de Discord: Controla compras semanales.

6 noviembre, 2025

To function optimally, AI browsers often require extensive access to personal resources, including email, calendar, and contact lists. Initial tests have revealed that while these agents exhibit moderate effectiveness for simpler tasks when given broad access, they frequently falter with more complex actions. Users may find the experience less about productivity and more akin to a novelty.

The significant concern lies within the realm of “prompt injection attacks,” vulnerabilities arising when malicious actors embed harmful commands within web pages. If an AI agent processes such content, it risks executing unintended commands, which could lead to unintentional data exposure or harmful actions, like unauthorized purchases.

Prompt injection attacks represent a growing threat associated with AI agents, with no comprehensive solution currently available. The launch of ChatGPT Atlas suggests an increase in user experimentation with AI browser agents, raising the stakes for security risks.

Brave, a browser company that prioritizes user privacy, has identified indirect prompt injection attacks as a systemic challenge in AI-powered browsers. This recognition calls for a reevaluation of security standards across the industry.

Experts including Shivan Sahib from Brave highlight both the opportunities AI browsers present and the inherent dangers of allowing them to act on a user’s behalf. OpenAI’s Chief Information Security Officer, Dane Stuckey, has also acknowledged these security challenges, noting that prompt injection remains an unresolved issue that adversaries are keen to exploit.

In response to these vulnerabilities, both OpenAI and Perplexity are working on preventive measures. OpenAI’s “logged out mode” restricts agents from accessing a user’s account, limiting potential data exposure, while Perplexity has developed real-time detection systems for prompt injection attacks. Despite these efforts, experts caution that no approach is infallible.

Security professionals indicate that large language models struggle to differentiate between legitimate instructions and harmful commands, creating ongoing challenges in thwarting prompt injection. Stephen Grobman from McAfee describes this evolving threat landscape as a “cat and mouse game,” where attack strategies continuously adapt to security measures.

Practical measures for users include employing unique passwords and multi-factor authentication to safeguard accounts. Additionally, users are advised to limit the access granted to these AI browsers and to isolate them from sensitive information, ensuring a layer of protection until these technologies mature.

While the information reflects the state of AI browsing tools as of October 2025, the rapid advancement of technology implies that ongoing updates and improvements in security practices are essential as we navigate this evolving digital landscape.

Gracias por leer Columna Digital, puedes seguirnos en Facebook, Twitter, Instagram o visitar nuestra página oficial. No olvides comentar sobre este articulo directamente en la parte inferior de esta página, tu comentario es muy importante para nuestra área de redacción y nuestros lectores.