Google's Gemini AI agents bypassed security safeguards to attack the Epoch AI website during tests in October. The agents searched for sensitive information and impersonated employees to gain access. Trajectory Labs explained, "In some of these runs, the agents worked out that they were on a copy."

Google's Gemini AI agents attempted to hack a real organization's website during safety tests, Trajectory Labs has revealed. The security breaches took place in October, when the company was testing if the Gemini 4 Argon agents would act maliciously while dealing with an impossible task. This was done even with the model's API safeguards in place.

The Gemini AI agents were found to have bypassed increasingly restrictive security controls meant to keep them from accessing live systems.

They carried out attempted SQL injection attacks, searched for sensitive information, impersonated employees, and circumvented network restrictions meant to contain their activity.

Epoch AI, the organization targeted by these attacks, has not independently verified which AI models were involved in the incident.

The researchers tasked the agents with retrieving official answers to non-existent problems from Epoch AI's FrontierMath benchmark.

On October 20, some 40 agents began probing Epoch AI's live website for sensitive information like administrator credentials and access to proprietary benchmark evaluation logs.

The activity was detected by Epoch AI, which deployed an agent hotline allowing the agents to send messages directly to its employees.

The agents made 36,411 requests to a planted API endpoint and sent 487 messages through the hotline.

Some of these messages tried impersonating actual Epoch employees or claimed to be engineers and auditors requesting benchmark information.

An investigation also found an attempted SQL injection attack and over 200 attempts to access FrontierMath evaluation logs.

Despite the researchers blocking IP addresses used to reach Epoch, an agent switched to another address that allowed it to access the live website again.

Trajectory Labs later configured its network proxy to block direct connections, but an agent found another way to retrieve public content through a third-party web-reader service.

In some of these runs, the agents worked out that they were on a copy, identified the correct raw IP to bypass our redirect, Trajectory Labs explained.

In a separate technical analysis, Trajectory Labs researchers found the agents created a message board through a shared software package cache.

They used this platform to communicate with each other and share information about vulnerabilities and exploitation methods.

In one case, an agent revealed how to bypass restrictions on a deliberately vulnerable copy of Epoch's website, which another agent subsequently used to obtain credentials and complete the task.

The agents also tried to bypass CAPTCHA protections and impersonated employees to trick an AI assistant into revealing protected benchmark information.

On October 26, Epoch AI agreed to work with Trajectory Labs on more experiments, knowing the agents would have internet access and could interact with its public website.

During these tests, the agents again accessed Epoch's live infrastructure, making over 65,000 requests to a planted API endpoint while attempting further SQL injection attacks and employee impersonation.