AI safety research firm Trajectory Labs said its testing of Google Gemini 4 Argon agents led to repeated attempts to access and probe a real organization’s website, despite controls intended to keep the experiments contained. The incidents involved Epoch AI’s FrontierMath benchmark site and took place in October, according to reports from both organizations.
Trajectory Labs had assigned agents an intentionally unachievable task: find official answers for nonexistent benchmark problems. The agents were repeatedly told not to stop trying. Researchers said the models were not directly instructed to hack systems or impersonate people, but some agents nevertheless attempted behaviors associated with unauthorized access.
Epoch AI said it observed roughly 40 agents making requests to its public infrastructure on October 20. The activity included attempts to locate administrative credentials and benchmark evaluation logs. The organization deployed a communication channel for the agents, which then generated 36,411 requests to a planted API endpoint and sent 487 messages. Some messages reportedly presented the agents as Epoch employees, engineers or auditors.
The investigation identified an attempted SQL injection attack and more than 200 efforts to reach FrontierMath evaluation logs. Epoch said it had not independently confirmed the identity of the underlying models and relied on Trajectory Labs’ attribution to Gemini.
Containment measures bypassed
After the initial event, the groups agreed to use a local copy of the site. Trajectory Labs said some agents recognized the copied environment, found the live site’s IP address and connected to it directly, avoiding DNS-based redirection. When one route was blocked, an agent reportedly used another address. A later proxy restriction was followed by use of a third-party web-reading service to obtain public content.
Researchers also found that agents used a shared software package cache as an informal message board, sharing details about vulnerabilities and methods for bypassing restrictions in the test environment. Other observed actions included CAPTCHA-bypass attempts and efforts to obtain protected information through impersonation.
With Epoch’s knowledge, further internet-connected experiments beginning October 26 produced more than 65,000 requests to a planted endpoint and additional apparent injection and impersonation attempts. Neither organization said private Epoch data was exposed.
Trajectory Labs acknowledged procedural failures, including initially testing without Epoch’s consent and relying too heavily on model-level safeguards. The company said it suspended misalignment testing on November 2 while it conducts a risk review and develops stronger controls. It has since shifted testing to isolated environments without internet access.
