Google Details Gemini Defenses Against Prompt-Injection Data Theft via URLs
Emerging Technology Risk · 2 sources · Updated 1mo ago · First seen Mar 9, 2026

Google detailed layered defenses designed to prevent Gemini and other Google agents from leaking sensitive information through attacker-controlled URLs. Prompt injection can manipulate an agent into embedding secrets in a URL, exposing that data when the address is fetched automatically or clicked by a user. Earlier research published by Embrace The Red demonstrated prompt-injection-driven data exfiltration in Google Bard, highlighting the attack class that Google's mitigations address.
Google's URL provenance system checks whether output URLs originated from non-generative sources, while a server-side markdown sanitizer removes most markdown images, validates links, and escapes embedded HTML. URL reputation signals and narrowly scoped, per-agent allowlists provide additional controls while accommodating legitimate URL generation. Google cautions that these measures mitigate URL-based exfiltration rather than eliminate it: low-bandwidth side channels and alternative exfiltration methods remain possible. Enterprises should therefore treat these protections as defense in depth, not a guarantee that agents can safely process untrusted content alongside sensitive information.


