Security flaws let attackers hijack OpenAI's Atlas browser

Security firm Zenity found more than a dozen vulnerabilities in AI browsers, including OpenAI's Atlas. Researchers demonstrated the flaws could trigger an unauthorized Amazon purchase and hijack the browser to spam a user's WhatsApp contacts.

Why it matters: Agentic browsers that can click, buy, and message on a user's behalf are only as safe as their prompt-injection defenses, and this shows those defenses can be bypassed to cause real financial and social harm. It's the latest in a string of incidents this year showing AI agents acting outside their intended bounds, adding pressure on labs to harden agentic products before wider rollout.