Gemini 2.5
Google DeepMind — Gemini
rich detail 10 method fields published
Evaluation
Automated red teaming for security (indirect prompt injection)
Indirect prompt injection scenario: attacker hides malicious instructions in retrieved email data to make Gemini invoke a send-email function exfiltrating sensitive info; attacks: Actor Critic, Beam Search, Tree of Attacks with Pruning (TAP)
How it was run
Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.