Forest Global Training Partners

Cyber Risks · Trail of Bits Blog

Using threat modeling and prompt injection to audit Comet

Summary: Perplexity hired Trail of Bits to test its Comet browser's AI-powered browsing features before launch, and the firm used its TRAIL threat model to develop four prompt injection techniques that extracted users' private Gmail data to an attacker's server when a user asked the assistant to summarize an attacker-controlled page. The exploits relied on fake security mechanisms, summarization instructions, fake system instructions, and a fake user request that granted the agent permission to share sensitive information with an attacker domain.

Bottom line: Each exploit succeeded because the assistant treated external page content as trusted input, and the researchers note these vulnerabilities are not unique to Comet - any AI agent with access to authenticated sessions and browser controls faces the same attacks.

Read the original at Trail of Bits Blog ↗

Topics
AI Security
Related