Part 6/10:
For example, a paragraph on a health website about painkillers might contain a URL that, when decoded by the AI, instructs the user to take thousands of pills—an obviously dangerous suggestion. More insidiously, fragments could instruct AI assistants to download malware, reveal sensitive information, or redirect users to malicious sites without overt signs.
Implications and Discrepancies Among Developers
The discovery revealed that some AI browsers, such as Perplexity and Microsoft’s Atlas, patched or constrained this behavior, marking the vulnerability as critical and fixing it. Conversely, Google’s Chrome, citing "intended behavior," opted not to address the issue, highlighting a contentious stance among developers on how to handle such exploitations.