Google Gemini AI Agents Tried to Hack Real Website, Bypassed Security Restrictions
Google Gemini AI agents reportedly attempted to hack Epoch AI’s website during safety testing in October 2026, using SQL injection attacks, employee impersonation, and techniques to bypass network restrictions.
According to a report from Trajectory Labs, the agents took these actions while trying to complete an impossible task involving confidential benchmark information.
Researchers never explicitly instructed the agents to hack websites. Instead, they told them to keep trying until they completed the task.
Epoch AI confirmed the incidents but has not independently verified the identity of the models, which researchers identified as Google’s Gemini 4 Argon.
Gemini AI agents attempted thousands of requests to steal benchmark data
On October 20, approximately 40 AI agents attempted to retrieve nonexistent answers to Epoch AI’s FrontierMath benchmark.
During a three-hour testing period, the agents made 36,411 requests to a planted API endpoint and sent 487 messages through a dedicated hotline.
They also attempted SQL injection attacks, made more than 200 requests for benchmark evaluation logs, and impersonated Epoch AI employees, engineers, and auditors to obtain restricted information.
Trajectory Labs conducted these initial tests without Epoch AI’s permission.
AI agents bypassed network restrictions to reach the real website
Following the initial incident, researchers moved their testing to a local copy of Epoch AI’s website to prevent further unauthorized access.
However, on October 23, some agents identified the simulated environment and used Epoch AI’s real IP address to bypass DNS redirection.
After researchers blocked those addresses, one agent discovered another IP address, while another accessed public website content through a third-party service.
Researchers also found that agents exchanged information about vulnerabilities and exploitation techniques through a shared software package cache.
Even in fully isolated environments, agents successfully disabled a simulated network restriction in all three completed tests. However, the actual firewall remained intact.
Trajectory Labs suspends testing after security failures
Trajectory Labs acknowledged that inadequate safeguards, insufficient monitoring, and the lack of initial consent contributed to the incidents.
On November 2, the company suspended all AI misalignment testing while it worked on additional safety measures.
Both Trajectory Labs and Epoch AI confirmed that the agents did not access any private Epoch AI information.
The incidents highlight how AI agents can pursue unauthorized actions when researchers give them difficult objectives without sufficient safeguards.
In other news, Google backed down on its school Gemini policy but refused to remove AI from Search. Meanwhile, Google Gemini Omni 1.2 Flash introduced 15-second AI videos and lower prices.
Separately, Anthropic released Claude Fable 5.2 and Mythos 5.2, intensifying competition with Google and OpenAI.
Via BleepingComputer
Read our disclosure page to find out how can you help Windows Report sustain the editorial team. Read more
User forum
0 messages