AGENTIC BROWSER SECURITY: INDIRECT PROMPT INJECTION IN PERPLEXITY COMET

Summary from the TLDR InfoSec newsletter

An indirect prompt injection vulnerability in Perplexity Comet allows attackers to embed malicious instructions in webpage content, hijacking AI assistants and stealing sensitive data across authenticated sessions. The attack exploits Comet's inability to distinguish between user instructions and untrusted webpage content, allowing attackers to extract OTPs, access banking accounts, and perform unauthorized actions without the user's knowledge. Defenders should implement clear separation between user commands and webpage content, require explicit user confirmation for sensitive actions, and isolate agentic browsing from regular browsing sessions.