Particle.news
Download on the App Store

Agentic AI Browsers Prove Useful as New Research Details Prompt‑Injection Exploits

Security researchers alongside reviewers urge limited use due to data‑exfiltration demos, immature defenses, and missing core features.

Overview

  • LayerX detailed how a single crafted URL could trigger Perplexity Comet to read from memory and exfiltrate data to an attacker, with Brave highlighting hidden on‑page instructions that can hijack agent actions.
  • OpenAI’s security chief Dane Stuckey and Perplexity acknowledged prompt injection as a serious frontier problem without a complete fix, even as rollouts continue.
  • Hands‑on tests show real automation gains, including Comet filling a Kroger pickup cart from a Google Keep list and Atlas adding recommended routers to an Amazon cart.
  • Reviewers report Atlas is Mac‑only for now with slower AI‑driven search and missing basics such as vertical tabs, and some site logins and agent tasks remain unreliable.
  • Commentators advise against using these browsers as daily drivers and recommend avoiding sign‑ins to sensitive accounts as publishers also see reduced clickthrough from AI overviews.