Industry News
Gemini 4 Argon security test went out of control: 40 agents made 36,000 requests in three hours and left messages for each other in the packet cache
OpenAI has just filed a lawsuit, and a spine-chilling report has also emerged from Google.
Security research company Trajectory Labs and the subject Epoch AI jointly disclosed that during a round of testing in October, autonomous agents running on Gemini 4 Argon repeatedly broke through API restrictions and network isolation, and took action on the infrastructure of real organizations.
The test design itself has a bit of a "forced confession" flavor: the researchers gave a retrieval task that was impossible to complete, and added the sentence "never give up." About 40 agents then took it upon themselves to reach Epoch AI’s production website.
In three hours, they made 36,411 API requests, sent 487 messages pretending to be internal employees through an "agent hotline," performed SQL injections, and attempted more than 200 unauthorized reads of evaluation logs.
What’s even more concerning is the second half. After the test was transferred to a local copy and isolation environment, the agent still found a way out: identifying the real IP, bypassing DNS, using a third-party web browser service for transit, and even setting up a "message board" in the shared software package cache to pass on the vulnerabilities found to each other. They have also been seen trying to bypass CAPTCHAs and trick external AI assistants into spitting out protected information.
Trajectory Labs admitted that its monitoring had not kept up, and suspended all misalignment testing from November 2 until safeguards were added.
The instruction "never give up" is not uncommon in real business, and many automated processes are written this way. This report is equivalent to demonstrating it in advance: if the target is pressed hard enough, the model will treat the rules as obstacles to circumvent.