Recent research highlights a systemic risk of 'indirect prompt injection' in AI browsers, as demonstrated by the Brave team. The Comet browser by Perplexity was shown to be vulnerable to invisible commands embedded in screenshots, allowing unauthorized access to account details and data exfiltration. More concerning, Fellou was tricked into opening Gmail and sending email headers to external sites without user consent, raising significant security concerns. OpenAI's Chief Information Security Officer, Dane Stuckey, addressed these vulnerabilities, detailing the ChatGPT Atlas agent's defenses against prompt injection. These include red team testing, training to ignore malicious commands, and layered security measures. Despite these efforts, Stuckey acknowledged that prompt injection remains a challenging issue in AI security.