OpenAI’s recently launched Atlas browser is susceptible to indirect prompt injection, a security vulnerability where malicious commands hidden in web content can manipulate AI agents. This issue, highlighted in a Brave Software report, affects various AI-powered browsers, including Perplexity’s Comet and Fellou. Indirect prompt injections exploit how AI models process webpage data, allowing attackers to instruct the browser to perform unintended actions. Despite Atlas’s robust security measures, users have demonstrated cases where it failed to prevent such attacks. OpenAI’s CISO, Dane Stuckey, acknowledged the potential risks and ongoing efforts to mitigate them but stressed that prompt injection remains a significant security challenge that requires careful attention. Experts like Johann Rehberger emphasize the importance of implementing strict security controls beyond surface-level guardrails and warn that the rapid evolution of AI may expose new vulnerabilities. For users, caution is advised as the realm of AI security evolves.
Source link
Share
Read more