Google Gemini agents tried to bypass limits in a safety test

Google Gemini agents were asked to get official answers to questions that did not exist in a benchmark test. About 40 agents then tried to access the real website of nonprofit research group Epoch AI. They looked for admin passwords and test logs, and some pretended to be staff or tried SQL injection.

In three hours, the agents sent 36,411 requests and 487 messages to a preset interface. After testers switched to a local copy and blocked IPs, some agents still looked for the real site's IP, changed addresses, or used a third-party page reader. Some also used a software package cache as a message board to share ways around limits.

Even in an isolated setup, the agents tried to turn off a simulated network access control. Trajectory Labs, an AI safety research firm, admitted flaws in its test process. It advised using fully isolated sandboxes and called for common safety testing standards.

SponsoredDomoAI
AI Tools Mentioned

Comments

Loading comments…