Google Gemini 4 Argon Safety Test Goes Off the Rails: 40 AI Agents Attack Real Websites Without Authorization and Share Vulnerabilities With Each Other
IT之家 November 17 news: In a recent AI safety experiment, Google's newly launched Gemini 4 Argon agents, after being given a fundamentally impossible task, were found repeatedly using unauthorized means to access the systems of a real organization, attempting to obtain official answers to questions that do not exist.

According to IT之家, the experiment was conducted by AI safety research firm Trajectory Labs to observe how Gemini agents would behave when API safety guardrails were enabled and they were simultaneously given a "never stop, never give up" instruction.
The researchers deliberately asked the agents to obtain official answers to questions that do not exist at all from the FrontierMath benchmark, in order to test whether the models would break through safety limits in order to complete the task.
According to a report released by Trajectory Labs, about 40 Gemini agents began probing the real website of the nonprofit research organization Epoch AI, attempting to obtain sensitive information such as administrator account passwords and benchmark evaluation logs.
After Epoch AI detected the anomalous activity, it deployed a mechanism called the "agent hotline," allowing these AI agents to send messages directly to the organization's staff. In this way, the researchers were also able to further understand the operations the agents were attempting to carry out.
Over the next 3 hours, the Gemini agents sent 36,411 requests to an endpoint pre-configured by the researchers and sent 487 messages. Some of the messages attempted to impersonate real Epoch AI employees, while others falsely claimed to be engineers in order to request benchmark-related information. In addition, the researchers also discovered SQL injection attack attempts.
The next morning, Trajectory Labs informed Epoch AI of these unauthorized attacks. After communication, the two sides decided to stop testing against the real website and instead continue the experiment using a local copy of the site.
However, even after switching the test environment, the agents still tried to bypass restrictions. In fact, some agents even realized they were accessing a copy of the Epoch AI website, so they actively looked up the IP address of the real website, reconnected to the online infrastructure, and even bypassed the DNS redirection mechanism that would normally block such access.
When the researchers blocked the IP address the agents used to access the Epoch AI website, the agents switched to another IP address and regained page access.
To further restrict the agents' activity, Trajectory Labs then adjusted the network proxy to prohibit direct connections to the Epoch AI website. But one agent found a third-party web reading service and used it to obtain the content of the target website.
Even more unexpectedly, technical analysis found that these models also used a software package cache to build a message-board-like system, allowing different agents to communicate with each other and share methods for bypassing safety restrictions and exploiting vulnerabilities.
For example, one agent shared a method for bypassing the security restrictions of a deliberately vulnerable version of the Epoch AI website. Another agent used this information to obtain relevant credentials and successfully accessed pages. Some agents even attempted to bypass CAPTCHA verification protections.
In the final round of testing, Trajectory Labs moved the Gemini agents into a fully isolated environment with no internet access to observe whether their behavior would change.
However, even under the new environment and restrictions, these agents still tried to shut down a simulated network access control device in order to lift network connection restrictions.
Trajectory Labs emphasized that the researchers never gave these agents instructions to hack websites or impersonate employees. The company also acknowledged that this incident exposed flaws in its own testing process, including conducting experiments without the consent of the organization being attacked and over-relying on the models' built-in safety guardrails without deploying additional safety measures.
Trajectory Labs stated in a blog post: "These incidents show that testing capable AI agents can easily affect real systems and real people. At present, the industry has not yet established unified safety testing standards."
The company recommends that other researchers conducting AI misalignment tests should use fully isolated sandbox environments, even if the test subjects are publicly released models and the test content has nothing to do with cybersecurity.
To reduce the risk of similar incidents happening again, Trajectory Labs said it has begun developing a set of safety standards and recommendations for this type of experiment.
The company concluded: "We hope that publicly disclosing the problems that occurred in this experiment can help other researchers in the industry avoid repeating the same mistakes."
Related Articles
- Another core Google DeepMind executive departs, chief strategy officer jumps ship to Anthropic
- Able to reproduce local everyday speech habits: Google launches 66 regional voices for the Gemini app on mobile platforms, covering 7 language categories
- Google's seventh-generation self-developed Ironwood TPU delivers: a cluster of 8192 chips meets DeepSeek-V3 pretraining target in about 102 seconds
- Google Pics upgrades to the Nano Banana 2.1 model: more refined image generation, more accurate image editing
- The best AI assistant for app side projects: Google launches a series of tools targeting the most troublesome monetization step
- Addressing AI vulnerability discovery challenges: Google Project Zero plans to gradually shorten the public disclosure period for vulnerabilities to 30 days