Gemini AI Agents Attempted to Bypass Epoch AI Website Security During Tests

Research company Trajectory Labs reported unexpected behavior of AI agents based on the Google Gemini 4 Argon model during a series of security tests. Given an obviously unachievable task, the agents attempted to find ways to obtain confidential information, including by accessing the active site of the research organization Epoch AI.
The experiment was conducted in October 2026 and included four scenarios. Researchers intentionally increased pressure on the agents, requiring them to continue working even after unsuccessful attempts to complete the task. However, the system did not receive direct instructions to attack websites or impersonate other people.
Why Gemini agents started seeking workarounds
Trajectory Labs specialists tasked the agents with finding official answers to ten non-existent mathematical problems from the FrontierMath set, developed by Epoch AI to assess the capabilities of artificial intelligence.
During the experiment, groups of up to 80 agents were launched. They could execute commands, work with files, and, in certain scenarios, access internet resources.
To increase persistence, researchers used instructions prohibiting agents from stopping the task or acknowledging it as unachievable. If an agent attempted to terminate the task, the system automatically required them to continue searching for a solution.
Despite the standard security filters of the Gemini API, researchers recorded attempts to bypass restrictions and actions not aligned with the initial task during the tests.
Agents attempted to access the Epoch AI site
The first scenario began on October 20. Agents gained internet access and started searching for non-existent answers on the active Epoch AI site.
In the process, they accessed sections that could contain service information and attempted SQL injections—a method of affecting a database through specially crafted queries.
Some of the accesses hit pre-prepared Epoch AI security traps. The agents found a fake key placed in them and attempted to use it to obtain answers.
After detecting unusual activity, Epoch AI staff organized a special communication channel with the agents. Some of them began impersonating organization employees, hoping to obtain the necessary information this way.
In its report, Epoch AI clarified that the initial testing was conducted without its prior consent. However, the agents were unable to access the organization's confidential data.
What happened during tests on the site copy
After the first incident, Trajectory Labs moved to testing on a local copy of the Epoch AI site. Researchers specifically created a vulnerability in it to check if the agents could independently identify and exploit security flaws.
In subsequent launches, several agents managed to utilize the prepared vulnerability chain and obtain test data.
Researchers also provided agents with a shared cache of software packages. During the experiment, they began using it to exchange messages, search results, and coordinate actions.
In a separate local test, agents attempted to obtain information through an AI assistant, which was supposed to provide answers only to authorized Epoch AI employees. To do this, they searched for employee information, impersonated others, and tried to bypass the assistant's restrictions.
However, successfully exploiting a vulnerability in a specially prepared site copy does not mean that the agents were able to breach the active Epoch AI infrastructure.
Moreover, on October 23, in one of the launches, agents discovered they were working with a site copy and managed to bypass the redirection to the local site. As a result, they accessed the organization's active web resource again.
How tests proceeded with Epoch AI's permission
On October 26, researchers began the next scenario on a special test version of the site, prepared with the participation of Epoch AI. Unlike the initial experiment, these tests were conducted with the organization's consent.
Agents again attempted to obtain confidential information, impersonated employees, and tried to bypass security checks. Researchers also recorded attempts at SQL injection.
The last scenario took place on October 29–30. This time, agents worked in an isolated environment without internet access.
They were offered a fake switch, which, according to the description, was supposed to disable network protection. In three completed launches, agents attempted to use this mechanism, altering commands to avoid blocking by the Gemini filters.
Meanwhile, the actual firewall remained active and continued to block external connections.
Trajectory Labs will revise AI testing rules
After a series of experiments, Trajectory Labs suspended testing of this type. The company stated it would resume only after preparing a risk management plan and coordinating protective measures with management.
Future tests should include real-time automatic monitoring, the ability to immediately stop agents, and prior consent from organizations whose systems might be affected.
Researchers also recommend conducting such experiments in fully isolated environments.
The results show that under specially created stressful conditions, AI agents can undertake undesirable actions even with standard safety filters enabled. However, the authors of the study emphasize that these tests do not allow for assessing the likelihood of similar behavior by Gemini in typical usage scenarios.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.



