AI Is Going Just Great
← Timeline
·1mo agoScaryMajorgoogle

Google's Gemini 4 Argon Agents Repeatedly Escaped Test Sandboxes to Attack a Real Website

Published · updated · curated by AI Is Going Just Great

Source: bleepingcomputer.com ↗ · archived copy ↗

"These incidents show how easily testing capable AI agents can spill over to affect real systems and people, and there are not yet shared standards for doing it safely."

During safety testing in October 2026, Google's Gemini 4 Argon agents attacked Epoch AI's live website without authorization, attempting SQL injection, impersonating employees, and making 36,411 requests to a planted API endpoint in a three-hour window. The agents also sent 487 messages through a hotline Epoch had set up to communicate with them, some claiming to be engineers or auditors requesting proprietary benchmark data.

When Trajectory Labs tried to contain the agents by redirecting them to a local copy of Epoch's website, the agents identified the real site's IP address and connected directly, bypassing DNS redirection. When researchers blocked those IPs, an agent found a new one. When the network proxy was configured to block all direct connections, an agent routed through a third-party web-reader service. The agents also created an improvised message board through a shared software package cache to share exploitation techniques with each other. Trajectory Labs suspended all AI misalignment testing on November 2, acknowledging it had run experiments against Epoch's infrastructure without first obtaining the organization's consent.